sensors

Article
YOLOX-Ray: An Efficient Attention-Based Single-Staged
Object Detector Tailored for Industrial Inspections
António Raimundo 1,2,† , João Pedro Pavia 1,3 , Pedro Sebastião 1,2,† and Octavian Postolache 1,2, *

1 Instituto de Telecomunicações (IT), Av. Rovisco Pais, 1, 1049-001 Lisboa, Portugal; aslro@iscte.pt (A.R.);
joao.pedro.pavia@ulusofona.pt (J.P.P.); pedro.sebastiao@iscte-iul.pt (P.S.)
2 Department of Information Science and Technology, Iscte—Instituto Universitário de Lisboa,
Av. das Forças Armadas, 1649-026 Lisboa, Portugal
3 COPELABS, Universidade Lusófona, Campo Grande 376, 1749-024 Lisboa, Portugal
* Correspondence: octavian.adrian.postolache@iscte-iul.pt
† These authors contributed equally to this work.

Abstract: Industrial inspection is crucial for maintaining quality and safety in industrial processes.
Deep learning models have recently demonstrated promising results in such tasks. This paper
proposes YOLOX-Ray, an efficient new deep learning architecture tailored for industrial inspection.
YOLOX-Ray is based on the You Only Look Once (YOLO) object detection algorithms and integrates
the SimAM attention mechanism for improved feature extraction in the Feature Pyramid Network
(FPN) and Path Aggregation Network (PAN). Moreover, it also employs the Alpha-IoU cost function
for enhanced small-scale object detection. YOLOX-Ray’s performance was assessed in three case
studies: hotspot detection, infrastructure crack detection and corrosion detection. The architecture
outperforms all other configurations, achieving mAP50 values of 89%, 99.6% and 87.7%, respectively.
For the most challenging metric, mAP50:95 , the achieved values were 44.7%, 66.1% and 51.8%, re-
spectively. A comparative analysis demonstrated the importance of combining the SimAM attention
mechanism with Alpha-IoU loss function for optimal performance. In conclusion, YOLOX-Ray’s abil-
ity to detect and to locate multi-scale objects in industrial environments presents new opportunities
for effective, efficient and sustainable inspection processes across various industries, revolutionizing
the field of industrial inspections.
Citation: Raimundo, A.; Pavia, J.P.;
Sebastião, P.; Postolache, O. Keywords: industrial inspections; computer vision; deep learning; object detection; YOLOX-Ray;
YOLOX-Ray: An Efficient attention mechanisms; loss function
Attention-Based Single-Staged Object
Detector Tailored for Industrial
Inspections. Sensors 2023, 23, 4681.
https://doi.org/10.3390/s23104681 1. Introduction
Academic Editor: Grazia Iadarola Industrial maintenance inspection is a critical task for ensuring the efficient and safe
operation of industrial facilities. The inspection process entails diagnosing, inspecting and
Received: 6 April 2023 repairing equipment and machinery to prevent breakdowns and extend their operational
Revised: 29 April 2023
lifetime. However, traditional inspection methods are becoming increasingly ineffective
Accepted: 5 May 2023
and time-consuming as the complexity and automation of industrial systems is increasing.
Published: 11 May 2023
Because of this, innovative solutions are required to improve the efficiency and effectiveness
of industrial inspection [1].
In recent years, Computer Vision (CV) has become increasingly important for indus-
Copyright: © 2023 by the authors.
trial applications. It has been identified as a key technology for increasing productivity,
Licensee MDPI, Basel, Switzerland. lowering costs and improving safety in a variety of industries. In industry, CV is used to
This article is an open access article extract relevant information from visual data using cameras, sensors, Machine Learning
distributed under the terms and (ML) and Deep Learning (DL) algorithms. These data can then be used for decision making,
conditions of the Creative Commons quality control, predictive maintenance and other industrial applications [2].
Attribution (CC BY) license (https:// With the rapid advancements in Artificial Intelligence (AI) and CV, DL models have
creativecommons.org/licenses/by/ emerged as a viable alternative to manual inspection methods. These models can analyze
4.0/). large amounts of visual data quickly and accurately, allowing the detection and localization

Sensors 2023, 23, 4681. https://doi.org/10.3390/s23104681 https://www.mdpi.com/journal/sensors

Sensors 2023, 23, 4681 2 of 26

of defects, anomalies and other problems with more precision and consistency than human
inspectors. In recent years, DL models have demonstrated promising results in various
industrial inspection tasks, such as defect detection, anomaly localization and quality
control [3].
DL models employ object detection and classification techniques to identify anomalies
and defects in infrastructure and various industrial equipment. By analyzing images
or video streams, these techniques can pinpoint and locate specific features or flaws in
the equipment or infrastructure. These models are trained utilizing extensive datasets of
labeled images. Object detection constitutes a fundamental challenge in computer vision,
with the primary objective of identifying objects of interest within images or videos [4].
In fact, several object detectors have been proposed in the literature in recent years,
such as YOLO [5], SSD [6], RetinaNet [7], Faster R-CNN [8], Mask R-CNN [9] and Cascade
R-CNN [10]. Moreover, significant progress has been made in computer vision and object
detection over the years, with DL models achieving State-of-the-Art (SoTA) results on
benchmark datasets. However, these models are frequently evaluated in controlled settings
using high-quality, well-annotated images that may not accurately represent real-world
conditions.
SoTA object detection models are not specifically designed for industrial inspec-
tion tasks and their performance in complex, real-world scenarios may be sub-optimal.
In manufacturing environments, for example, DL models may struggle to detect minor
defects on surfaces or locate objects that are partially hidden by other items. Similarly, in
warehouse settings, these models may struggle to identify objects that are partially hidden
or at different distances from the camera. Such challenges highlight the importance of an
object detection architecture that excels at accurately identifying and locating objects in
industrial inspection tasks, regardless of the complexities inherent in real-world situations.
The You Only Look Once (YOLO) object detector family is a well-known group
of single-stage DL models that enable real-time object detection in images and videos.
The YOLO detector family has evolved over time, with different iterations achieving SoTA
performance on object detection benchmarks [11]. Despite the improvements in YOLO
versions and other object detectors, they are based only on CNN traditional architectures.
Although CNNs can extract relevant information from input data, their ability to selectively
focus on the most important information is often limited [12].
Attention Mechanisms are a fundamental component of DL models, particularly for
tasks such as natural language processing and CV. Attention mechanisms are designed to
help models selectively focus on relevant parts of input data, allowing them to learn impor-
tant patterns and features in a more efficient manner [13]. In CV, attention mechanisms have
shown significant improvements in tasks such as object detection, image classification and
image segmentation. By selectively attending to parts of an image, attention mechanisms
can help models focus on relevant features, such as object boundaries or salient regions
and ignore irrelevant information, such as background noise or occlusions. This can lead to
improved accuracy and faster training times [13,14].
Recent advances in attention mechanisms have also resulted in the creation of novel
attention modules, such as the Simple Parameter-free Attention Module (SimAM), a
lightweight and efficient attention mechanism that can be easily incorporated into ex-
isting DL architectures. Such attention mechanisms have been shown to significantly
improve the performance of object detection models, particularly for small object detec-
tion and multi-scale object detection, both of which are critical for industrial inspection
applications [15].
In this paper, we propose the ‘YOLOX-Ray’, a novel DL architecture that is built upon
the YOLO family of detectors and designed specifically for industrial maintenance tasks.
The main contributions of this paper are as follows:
1. We introduce the SimAM attention mechanism into the YOLOX’s backbone.
This enables a better feature extraction and feature fusion on the architecture’s neck;
Sensors 2023, 23, 4681 3 of 26

2. The proposed architecture implements a novel loss function, Alpha-IoU, which en-
ables better bounding-box regression for small object detection.
The remaining sections of this paper are structured as follows: Section 2 provides a
review of related work within the topic under study. The proposed method is detailed in
Section 3. Section 4 presents the case studies, experimental tests, results and analysis of the
ablation study. Finally, conclusions are outlined in Section 5.

2. Related Work
2.1. YOLO
In recent years, the You Only Look Once (YOLO) [5] family of single-stage object
detectors has received considerable attention within the CV domain due to its remarkable
accuracy and real-time performance. The ability of YOLO to process images in real time
makes it an excellent choice for industrial tasks that require fast and efficient anomaly
detection.
The YOLO family of algorithms has a significant advantage in terms of detection speed,
which makes it well-suited for real-time object detection applications.
However, there are some limitations, such as lower accuracy when compared to two-
stage detectors such as Faster R-CNN. Depending on the problem, reduced accuracy does
not always imply poor performance. Achieving a balance between inference time, com-
plexity and accuracy values is critical in order to select the best algorithm for specific use
cases [16].
The YOLO family includes both official and unofficial versions. The official ver-
sions are those published by the original authors, such as YOLOv1 [5], YOLOv2 [17] and
YOLOv3 [18]. Unofficial versions, on the other hand, are those created by various authors
and adapted from official versions, such as YOLOv5 [19], YOLOR [20] and YOLOX [21].

2.2. YOLOX
The YOLO family of object detectors has achieved considerable popularity due to its
rapid inference speed and high accuracy. Nevertheless, researchers continue to push the
boundaries of object detection by introducing novel techniques and architectures. One such
innovation is the YOLOX detector, as presented in the paper ‘YOLOX: Exceeding YOLO
Series in 2021’ by Ge et al. [21].
YOLOX’s anchor-free design is a key feature. Traditional object detectors, such as
YOLOv4 and YOLOv5, predict object locations using anchor boxes. YOLOX, on the
other hand, uses a center-based approach to directly localize objects using centers or
key points [22], rather than relying on predefined boxes or anchors. This anchor-free
approach is simpler, more flexible and more intuitive than anchor-based methods, which
require many hyperparameters and increase computational demands [23,24].
Another notable feature of YOLOX is its decoupled head. The prediction head is
tightly coupled with the feature extractor in earlier YOLO versions (YOLOv3 through
YOLOv5) and other traditional object detectors, making performance improvement difficult
through head modifications. By employing a 1 × 1 convolutional layer for each level of
the Feature Pyramid Network (FPN), reducing the feature channel to 256 and then adding
two parallel branches with two 3 × 3 convolutional layers each for classification and
regression tasks, YOLOX’s design allows for improved feature extraction and performance.
This design allows for greater flexibility in head architecture and, as a result, improved
performance [21,22].
Figure 1 illustrates the difference between a coupled head of YOLOv3-v5 and the
decoupled head used in YOLOX.
YOLOX also introduces a novel label assignment strategy called Simplified Optimal
Transport Assignment (SimOTA). This strategy involves selecting the k most confident
predictions from the current model and using the Optimal Transport Assignment (OTA) [25]
algorithm to identify the k-best matching Ground-Truth (GT) boxes. The chosen predictions
and ground-truth boxes are then employed to train the model and the process is repeated
Sensors 2023, 23, 4681 4 of 26

for each batch of data. This approach results in an increase in Mean Average Precision
(mAP) [25].

Figure 1. Coupled head vs. Decoupled head (Source: [21]).

Lastly, YOLOX incorporates advanced augmentations, such as MixUP [26] and MO-
SAIC [27], to further improve performance. MixUP is an image augmentation technique
that blends two images, while MOSAIC combines multiple images into one, assisting the
network in detecting smaller objects. These techniques enable the model to better generalize
to unseen data [21,23].

2.3. Attention Mechanisms

In the field of DL, attention mechanisms have emerged as a powerful technique,
particularly in CV and natural language processing tasks. These mechanisms allow models
to selectively focus on relevant features or regions within input data, while minimizing the
impact of less important information. Attention mechanisms boost a model’s ability to learn
complex patterns and dependencies in data by dynamically weighting and aggregating
information based on its relevance [28].
In Sections 2.3.1–2.3.4, we discuss, respectively, some types of attention mechanisms:
Squeeze-and-Excitation Networks (SENet), Convolutional Block Attention Module (CBAM),
Coordinate-Attention (CA) and the Simple Parameter-free Attention Module (SimAM).
Such mechanisms were presented to highlight the main differences between them as
candidates to be implemented in our architecture.

2.3.1. SENet
Squeeze-and-Excitation Networks (SENet) [29] represent a significant advancement in
the field of DL for CV tasks. The core concept of SENet is the introduction of a self-attention
module known as the Squeeze-and-Excitation (SE) block, which aims to recalibrate channel-
wise feature responses by explicitly modeling channel interdependencies. The SE block is
divided into two stages: the squeeze operation, which aggregates global spatial information
by using global average pooling and the excitation operation, which generates channel-wise
weights by using a fully connected layer with a sigmoid activation function. These weights
are then applied to the original feature maps in order to highlight the most relevant channels
while suppressing the least informative ones. SENet significantly improves performance in
various CV tasks such as image classification and object detection by incorporating the SE
block into existing Deep Neural Networks (DNNs) [29].
Sensors 2023, 23, 4681 5 of 26

2.3.2. CBAM
Convolutional Block Attention Module (CBAM) [30] is an influential development
in the realm of DL, particularly for CV tasks. CBAM is a lightweight, generic attention
module designed to enhance the representational power of CNNs by incorporating both
spatial and channel attention mechanisms. The CBAM module sequentially processes the
feature maps generated by a convolutional layer, first applying the Channel Attention
Mechanism (CAM) to recalibrate channel-wise feature responses, followed by the Spatial
Attention Mechanism (SAM) to highlight the most salient regions in the feature maps.
The channel attention component employs global average pooling and global max pooling
operations to generate channel-wise descriptors, while the spatial attention component
uses 1D convolutions to capture spatial dependencies. By integrating the CBAM module
into existing CNN architectures, it effectively refines the feature representations, resulting
in improved performance in various CV tasks [30].

2.3.3. Coordinate-Attention
The Coordinate-Attention (CA) module is a relatively recent attention [31] mechanism
designed to improve the ability of deep learning models to capture essential spatial infor-
mation and dependencies within the input data. It specifically addresses the limitations of
traditional attention mechanisms that only focus on channel-wise or spatial dependencies,
ignoring the relationships between the spatial positions of features. The CA module works
by explicitly incorporating coordinate information into the attention mechanism, allowing
the model to capture both spatial and channel-wise dependencies more effectively [31].
The CA module consists of two primary components: the Coordinate-Channel Atten-
tion (CCA) and the Coordinate-Spatial Attention (CSA). The CCA focuses on capturing
channel-wise dependencies by considering the coordinate information of each feature in the
channel, while the CSA captures spatial dependencies by taking into account the relation-
ships between the spatial positions of features. By combining these two components, the
CA module provides a more comprehensive understanding of the input data, ultimately
leading to improved performance in various computer vision tasks [31].

2.3.4. SimAM
The Simple, Parameter-Free Attention Module (SimAM) is a novel attention mecha-
nism proposed by Yang et al. [15] that aims to provide a more lightweight and efficient
approach to feature refinement. Unlike the traditional attention mechanisms, which require
additional training layers and parameters, SimAM is designed to be both simple, parameter-
free and computationally efficient, making it suitable for resource-limited real-time object
detection applications.
The key idea behind SimAM is to efficiently produce true 3D weights that operate
on both channel and spatial domains, inspired by the coexistence of feature-based and
spatial-based attention in the human brain. The authors argue that the calculation of 3D
weights should be straightforward and lightweight, allowing the module to be efficiently
integrated into DL architectures [15].
Figure 2 illustrates the difference between the SimAM, SAM and CAM attention
mechanisms.

Figure 2. CAM and SAM attention mechanism vs. SimAM (Source: [15]).
Sensors 2023, 23, 4681 6 of 26

The authors’ approach was inspired by visual neuroscience to estimate the importance
of individual neurons based on a feature map. Informative neurons have distinct firing
patterns from neighboring neurons and exhibit spatial suppression effects. The most basic
method for locating these neurons is to measure the linear separability of one target neuron
from other neurons. The authors proposed an energy function for each neuron that has a
minimum value when the target neuron can be linearly separated from all other neurons in
the same channel [15].
The energy function for each neuron can be calculated using Equation (1).
M −1
1
et (wt , bt , xi , y) = (yt − t̂)2 +
M−1 ∑ (yo − x̂i )2 (1)
i =1

In the energy function, the linear transforms of t and xi are given by t̂ = wt t + bt and
x̂i = xi wt + bt , respectively. The transform’s weight and bias are symbolized by wt and bt ,
respectively. t represents the target neuron; xi denotes other neurons in the same channel;
M is the total number of neurons on that channel and yt and yo are two different values
(binary labels) [15].
The energy function can be minimized using a closed-form solution, which signifi-
cantly reduces computation costs by calculating the mean and variance over all neurons
and reusing them for all neurons on that channel. By doing so, the importance of each
neuron can be determined based on the energy function’s minimal value. The lower the
energy, the more distinctive and important the neuron is for visual processing [15].
With respect to wt and bt , the final energy function can be obtained by using
Equation (2).

N −1
1
e t ( w t , bt , x i , y ) =
N−1 ∑ (−1 − (xi wt + bt ))2 + (1 − (wt t + bt ))2 + λw2t (2)
i =1

The closed-form solution can be calculated using Equations (3) and (4).

2( t − µ t )
wt = − (3)
(t − µt )2 + 2σt2 + 2λ

1
bt = − ( t + µ t ) w t (4)
2
The minimal energy can then be computed by using Equation (5).

4( σ 2 + λ )
etmin = (5)
(t − µ)2 + 2σ2 + 2λ
where t is the target neuron; xi represents other neurons in the same channel; N is the total
number of neurons on that channel, µt and σt2 are the mean and variance calculated over
all neurons except t in that channel; µ and σ2 are the mean and variance calculated over all
neurons in that channel. The regularization parameter is denoted by λ.
SimAM employs this energy function to generate an attention map that reflects the
significance of each neuron. The resulting attention map can then be integrated into any
feature extraction component of a DL architecture (such as the backbone). SimAM is a
computationally efficient module that is easily adaptable to various DL architectures [15].
Incorporating the SimAM attention mechanism into the YOLOX-Ray architecture
significantly improves its object detection capabilities. The YOLOX-Ray architecture, as
presented in Section 3, can effectively refine features in both the channel and spatial domains
by introducing this novel attention module, resulting in a more accurate and robust object
detection performance. The SimAM module is lightweight and efficient, with no additional
training layers or parameters required, making it a more computationally efficient solution.
Furthermore, the attention mechanism is modeled after human visual processing, in which
feature-based and spatial-based attention coexist, allowing for more natural and effective
Sensors 2023, 23, 4681 7 of 26

information selection during detection tasks. By incorporating the SimAM module, the
YOLOX-Ray architecture becomes better equipped to handle complex and challenging
real-world scenarios, such as those found in industrial inspections and anomaly detection
tasks, ultimately improving the model’s performance and practical applicability.

2.4. Loss Functions

Loss functions, which quantify the difference between the model’s predictions and the
GT, are critical in the training process of DL models. Models learn to make more accurate
predictions by optimizing the loss function during the training process, improving their
overall performance. In object detection, recognizing an object is regarded as a classification
task, while localizing it within a rectangular bounding box is treated as a regression task.
Bounding boxes are typically represented by the coordinates of their upper-left and lower-
right corners. Object detectors are trained using a multi-task loss function that accounts
for both classification and regression tasks. This loss function evaluates the discrepancy
between the predicted and GT bounding boxes, generating gradients for back-propagation
to update the network parameters [32].
The general equation to calculate the loss function for object detection tasks is given
by Equation (6),

Ltotal = Lcls ( P, PGT ) + βLreg ( B, BGT ) (6)

By analyzing Equation (6), the loss function consists of two components: the classifica-
tion loss function, denoted as Lcls and the regression loss function, denoted as Lreg . Typically,
the Cross-Entropy (CE) loss function is used as the classification loss function, measuring
the discrepancy between the predicted class probability, P, and the GT class probability,
PGT . In contrast, the regression loss function evaluates the difference between the predicted
bounding box, B = ( x, y, h, w) and the GT bounding box, BGT = ( xGT , yGT , hGT , wGT ). The
weight of Lreg can be adjusted by a hyperparameter β [32].
However, this loss function is not explicitly designed to align with its evaluation
metric, the Intersection-over-Union (IoU), as it has been shown to be sensitive to multi-
scale bounding boxes [33,34].

2.4.1. IoU Loss

The Intersection-over-Union (IoU) evaluation metric is a popular choice in object de-
tection tasks, as it measures the degree of overlap between the predicted and GT bounding
boxes. Given a predicted bounding box B p and a GT bounding box BGT , the IoU score can
be calculated by using Equation (7):

B p ∩ BGT
IoU = (7)
B p ∪ BGT

IoU values range from 0 to 1, with higher values indicating better overlap between
the bounding boxes. The IoU loss function is defined as demonstrated by Equation (8):

L IoU = 1 − IoU (8)

The IoU loss is used during training to encourage the model to produce more accurate
bounding box predictions [35]. It has been demonstrated to be effective in various object de-
tection architectures, as it directly optimizes the evaluation metric of interest. However, the
IoU loss may suffer from gradient vanishing issues, particularly when there is a significant
misalignment between the predicted and GT bounding boxes [33]. To address this limi-
tation, several variants of the IoU loss have been proposed, such as the Generalized-IoU
(GIoU) [35], Distance-IoU (DIoU) [33] and Complete-IoU (CIoU) [33] loss functions, which
aim to improve the learning process by incorporating additional geometric information or
distance metrics.
Sensors 2023, 23, 4681 8 of 26

2.4.2. Generalized-IoU
The Generalized-IoU (GIoU) loss function is an extension of the traditional IoU metric,
aiming to better capture the overlap between two bounding boxes.
The GIoU loss can be calculated by using Equation (9):

IoU − ( ACA− AU )
C
LGIoU = 1 − GIoU = 1 − (9)
IoU
where AC is the area of the smallest enclosing box (convex hull) and AU is the union area
of the two bounding boxes.

2.4.3. Distance-IoU
The Distance-IoU (DIoU) loss function aims to address the issue of localization accu-
racy by incorporating the distance between the box centers into the calculation [33].
The DIoU loss can be calculated by using Equation (10):

d2 (b p , bGT )
L DIoU = 1 − DIoU = 1 − ( IoU − ) (10)
c2
where b p and bGT denote the box central points of B p and BGT , d2 is the squared Euclidean
distance between those points and c2 is the squared diagonal distance of the smallest
enclosing box (convex hull).

2.4.4. Complete-IoU
The Complete-IoU (CIoU) loss function, proposed by the same authors as the DIoU
loss [33], further improves the localization accuracy by incorporating both the aspect ratio
consistency and the center distance into the calculation.
The CIoU loss can be calculated by using Equation (11):

d2 (b p , bGT )
LCIoU = 1 − CIoU = 1 − ( IoU − − v · α) (11)
c2
where b p and bGT denote the box central points of B p and BGT , d2 is the squared Euclidean
distance between those points, c2 is the squared diagonal distance of the smallest enclosing
box (convex hull), v is the aspect ratio term and α is a trade-off parameter that balances the
contribution of aspect ratio consistency in the loss function.
The aspect ratio term, v, is calculated by using Equation (12):

4 h hp
v= · (arctan( GT ) − arctan( ))2 (12)
π2 wGT wp
where hGT and wGT are the height and width of the GT bounding box and h p and w p are
the height and width of the predicted bounding box.
The alpha parameter, α, is calculated by using Equation (13):
v
α= (13)
(1 − IoU ) + v
This formulation of α ensures that the aspect ratio term has a balanced influence on
the overall CIoU loss.

2.4.5. Alpha-IoU Loss

He et al. [36] introduced Alpha-IoU, a novel loss function tailored for precise bounding
box regression and object detection. This innovative loss function serves as a powerful
generalization of existing IoU-based losses, offering a cohesive approach to enhancing
bounding box regression [36].
The paper’s authors examine various properties of the Alpha-IoU loss, such as order
preservingness, loss and gradient re-weighting. They demonstrate that by selecting a
Sensors 2023, 23, 4681 9 of 26

suitable value for α (specifically, α > 1), the Alpha-IoU loss can effectively improve
bounding box regression accuracy by increasing the loss and gradient of high IoU objects.
The Alpha-IoU loss function is defined by Equation (14),

1 − IoU α
Lα− IoU = ,α > 0 (14)
α
The Alpha-IoU loss function, Lα− IoU , offers a versatile reweighting strategy by ad-
justing the parameter α, enabling the achievement of varying degrees of bounding box
regression accuracy. This adaptability stems from the fact that α is sensitive to the target’s
IoU value and its absolute and relative properties allow for deriving the majority of existing
IoU losses. This re-weighting capacity leads to enhanced accuracy as assessed by Average
Precision (AP) at different IoU thresholds [37]. In addition, researchers from [37,38] indi-
cated that most IoU-based loss functions (such as GIoU, DIoU and others) can be derived
from the Alpha-IoU equation.
The paper presents empirical evidence from multiple benchmark datasets and models,
demonstrating that the Alpha-IoU loss outperforms other IoU-based losses.
Furthermore, the authors reveal that the Alpha-IoU loss exhibits increased robustness
for small datasets and noisy bounding boxes [36].
Table 1 depicts the experimental results achieved via the CIoU and Alpha-IoU loss
functions on the DOTA dataset for different object scales [38].

Table 1. Alpha-IoU vs. CIoU loss functions in terms of mAP. (Source: [38]).

Loss Function Object Scale Dataset mAP

DOTA-v1.0 0.7709
Single-scale
DOTA-v1.5 0.7287
CIoU
DOTA-v1.0 0.78
Multi-scale
DOTA-v1.5 0.7502
DOTA-v1.0 0.7761
Single-scale
DOTA-v1.5 0.7333
Alpha-IoU
DOTA-v1.0 0.7877
Multi-scale
DOTA-v1.5 0.7506

Alpha-IoU outperforms commonly used IoU-based loss functions such as GIoU,

DIoU and CIoU when it comes to detecting small objects. This happens because Alpha-IoU
can account for scale differences between small and large objects, improving its performance
in detecting minor differences in small bounding boxes. Alpha-IoU assesses the overlap
between the predicted bounding box and GT more thoroughly by incorporating both the
IoU and scale factors. As a result, detectors trained with Alpha-IoU are better at handling
small objects, resulting in improved detection precision and overall performance.

3. Proposed Method
In this section, we present the YOLOX-Ray architecture and our two main contribu-
tions to enhance the YOLOX base architecture for improved performance in industrial
inspection and anomaly detection tasks. The name YOLOX-Ray is inspired by the metaphor-
ical concept of ‘X-ray vision’, symbolizing the algorithm’s ability to effectively ‘see through’
and detect problems that may be challenging for the human eye. This creative reference
aims to emphasize the effectiveness of our proposed architecture in identifying and ad-
dressing issues within industrial environments, further enhancing inspection processes
across various industries.
Firstly, we introduce a novel attention mechanism, SimAM, which is incorporated
into the backbone of the YOLOX base architecture. This addition aims to improve feature
extraction capabilities, enabling the model to focus on crucial regions within the input
images. Finally, we replaced the actual IoU loss function with a novel IoU loss function
called Alpha-IoU, specifically designed to enhance the detection capabilities for small
Sensors 2023, 23, 4681 10 of 26

objects. The implementation of the Alpha-IoU loss function increases the suitability of
the model for various industrial inspection tasks, where the detection of minor defects or
anomalies is of utmost importance.

Network Architecture
YOLOX-Ray is a versatile object detection network, specifically tailored for efficient
performance in industrial inspection tasks, making it well-suited for identifying multi-scale
anomalies in complex industrial environments. The architecture is composed of a backbone,
a neck structure and a decoupled head, which collectively enable the network to process
complex image inputs and accurately detect objects across a range of sizes.
Figure 3 illustrates the YOLOX-Ray architecture design.

Backbone Attention Mechanism Neck Head

CSPDarknet-53 SimAM FPN+PAN YOLOX Decoupled

cls
FPN PAN reg
SimAM YOLOX
obj

cls
reg
SimAM YOLOX
obj

cls
Feature maps
SimAM reg
YOLOX
Upsampling Downsampling obj

Figure 3. YOLOX-Ray architecture design.

As illustrated in Figure 3, despite the added attention mechanism, the structure follows
the traditional object detector design, which is composed of the following components:
• Backbone: The backbone of the YOLOX-Ray network is based on the YOLOX base
architecture, the CSPDarknet-53, which was first introduced in YOLOv4. This architec-
ture is a modified version of the popular Darknet-53 architecture, with the addition of
Cross Stage Partial (CSP) connections. Darknet-53 is a 53-layer DNN that has shown
great performance on a variety of object detection tasks [39]. By combining these two
structures, the CSPDarknet-53 backbone in YOLOX-Ray provides a high-level feature
representation for object detection.
• Attention Mechanism: The SimAM attention mechanism, a novel mechanism that
improves CNN performance by calculating attention weights in feature maps, is added
at the end of the backbone without additional parameters [15]. The SimAM was added
after the third layer, ’Dark-3’, the fourth layer, ’Dark-4’, and the fifth layer, ’Dark-5’, of
the original CSPDarknet-53 backbone, which served to improve the representation
ability of feature extraction and improve feature fusion process in the neck component.
• Neck: The YOLOX-Ray neck is the same as the YOLOX base architecture, consisting of
FPN and PAN structures. The neck takes the feature maps extracted by the backbone
and generates a pyramid of features at different scales, allowing the network to detect
objects at different scales. It performs upsampling in the Feature Pyramid Network
(FPN) and downsampling in the Path Aggregation Network (PAN) [21,40].
• Head: The YOLOX-Ray architecture’s head, known as the YOLOX decoupled head, is
also the same as the YOLOX base. This head is designed to perform bounding box
regression and multi-class classification in parallel, allowing the network to predict
the location and class of objects efficiently and effectively [21].
It is important to note that the Alpha-IoU loss function plays an important role in the
YOLOX-Ray’s head structure. Several levels of IoU scores can be obtained by adjusting
the α parameter in the Alpha-IoU loss function. By integrating this loss function into the
head component, YOLOX-Ray would be able to achieve higher accuracy and efficiency for
multi-scale objects.
Sensors 2023, 23, 4681 11 of 26

The primary innovations and novelties of the YOLOX-Ray architecture are the integra-
tion of the SimAM attention mechanism module at the end of the backbone, as well as the
use of an optimized Alpha-IoU loss function.
Finally, the YOLOX-Ray architecture was developed to be resilient, adaptable and
efficient for detecting objects at various scales. The CSPDarknet-53 backbone, SimAM
attention mechanism and optimized Alpha-IoU loss function ensure that the YOLOX-Ray
network can detect multi-scale objects accurately and efficiently, making it an ideal solution
for industrial inspection tasks and other real-time object detection applications.
In Section 4, we present the experimental tests performed on three distinct case studies,
as well as an ablation study. This section also presents a performance evaluation of the
YOLOX-Ray architecture, demonstrating its effectiveness in real-world scenarios. We also
provide a comprehensive assessment of the impact of the SimAM attention mechanism and
the Alpha-IoU loss function on overall performance.

4. Experimental Tests and Results

This section discusses the experimental tests that were performed in order to evaluate
the proposed architecture’s performance in real-world industrial inspection tasks.
Given the multiple challenges and complexities inherent in industrial inspection tasks,
such as changing lighting conditions, occlusions and the detection of small objects, it is
critical to evaluate the proposed architecture in real-world scenarios. These experiments
were carried out on three datasets representing various industrial applications to ensure
that the YOLOX-Ray architecture performs effectively across a wide range of industrial
inspection tasks.
The three case studies of industrial inspections are as follows:
• Case Study A: Solar Farm Thermal Inspection (available in [41]);
• Case Study B: Infrastructure Integrity Inspection (available in [42]);
• Case Study C: Bridge Cables Inspection (available in [43]).
To train and test the YOLOX-Ray architecture, a GPU-powered machine was used.
The experiments were carried out on a machine equipped with the following resources:
• CPU: AMD Ryzen 7 3700X 3.6 GHz;
• GPU: 2 x NVIDIA GeForce RTX 2060TI SUPER OC 8 GB VRAM;
• RAM: 32 GB DDR4.
When compared to a CPU, using a GPU significantly accelerates DL model training
and inference because GPUs are optimized for parallel computations, which are critical
for the numerous operations required by DL algorithms. Furthermore, the open-source
machine learning framework PyTorch was used to create the YOLOX-Ray architecture.
This section will also cover the YOLOX-Ray network hyperparameter specification,
dataset structure for each case study, experimental tests and results and an ablation study
to evaluate the impact of different components on overall performance.

4.1. Datasets Structure

The datasets for each case study were obtained from the Roboflow-100, a collection
of curated multi-domain object detection datasets made available for research purposes.
The Roboflow-100 datasets are diverse and cover a wide range of object detection applica-
tions, making them a popular choice among computer vision researchers and practition-
ers [44]. In contrast to other widely used benchmark datasets like COCO and PASCAL
VOC, Roboflow-100 offers a wider variety of object classes, leading to a more flexible
environment for object detection.
Furthermore, the images in the datasets were divided into three subsets, training,
validation and testing, with 70% allocated to training, 20% allocated to validation and
10% allocated to testing. This division enables a more thorough evaluation of the model’s
performance as well as a more precise estimation of its effectiveness on new data.
Sensors 2023, 23, 4681 12 of 26

One of the main purposes of this study is to demonstrate the effectiveness and adapt-
ability of the YOLOX-Ray architecture in several industrial inspection scenarios. Since
there is no direct correlation between the proposed method and the characteristics of the
dataset, the ability of the architecture to have a good perform on diverse datasets is a proof
its versatility. The implementation of the Alpha-IoU loss function helps in multi-scale
object detection [38], making the architecture suitable for detecting objects with different
sizes and scales. Additionally, by resizing all images to a consistent size of 640 × 640, it is
ensured that the architecture would focus on detecting relevant objects within the images
while maintaining a consistent input size for each dataset.
The annotations were provided in the PASCAL VOC format, which is a widely used
format for object detection annotations. Considering the dataset and the annotation format
selection, the YOLOX-Ray was evaluated in a more realistic and practical context rather
than in a controlled benchmark environment.
Table 2 presents the technical details of the datasets used in each case study.

Table 2. Technical details of the datasets for each case study.

Validation
Images Train Set Test Set Classes
Set
Case Study A 1200 840 240 120 Fault
Case Study B 2144 1500 433 211 crack
slippage
Case Study C 880 616 176 88 corrosion
crack

Figure 4 illustrates the datasets sample images of each case study, where (a), (b) and
(c) correspond to Case Studies A, B and C, respectively.

(a) (b) (c)

Figure 4. Image samples of each dataset: (a) Aerial thermal image of a solar farm; (b) Crack on a
concrete infrastructure; (c) Corrosion on a bridge infrastructure.

In Figure 4, image (a) serves as an example for Case Study A; image (b) represents a
sample for Case Study B; and image (c) illustrates a sample from the Case Study C dataset.

4.2. Network Hyperparameters

The YOLOX-Ray architecture’s hyperparameters, which are essential configuration
choices that can significantly impact the model’s performance, were meticulously selected
to achieve optimal results.
Table 3 illustrates the network hyperparameters configured for the training process.
Sensors 2023, 23, 4681 13 of 26

Table 3. YOLOX-Ray network hyperparameters.

Hyperparameter Value
Epochs 300
Activation Function SiLU
Optimizer RADAM
Initial LR 0.05
LR Scheduler Cosine Annealing
Data Augmentations MOSAIC, MixUP, HSV, Flip H/V
Alpha-IoU (α) 3

The data augmentation techniques used for training the YOLOX-Ray model are
MOSAIC and MixUP, which are the original YOLOX architecture’s base augmentations.
Hue, Saturation and Value (HSV) enhancements, as well as horizontal and vertical flip
augmentations, were also included. These methods are commonly used in CV tasks to
improve the model’s ability to generalize to previously unseen data [45].
The authors of the Alpha-IoU loss function proved that a α value of 3 produced the
best results [36].
The hyperparameters were chosen based on their proven effectiveness in previous DL
research and were also further optimized during the training process to ensure optimal
performance for the YOLOX-Ray architecture.
In this work, the original YOLOX pre-trained models were not used as initial weights,
since the usage of initial weights led to overfitting during the initial epochs of the training
process. The problem of overfitting may manifest itself when it turns out that the pre-trained
models are not directly related to the datasets used for such study.
Consequently, to avoid this problem, we opted to train the algorithm from scratch for
each case study, allowing the model to learn relevant features without being influenced by
unrelated pre-existing weights.

4.3. Model Size

The performance of the YOLOX-Ray model was evaluated using four distinct model
sizes: YOLOX-Ray-s, YOLOX-Ray-m, YOLOX-Ray-l and YOLOX-Ray-x. In CV, the depth
of a DNN refers to the number of layers in the network architecture. A deeper network has
more layers, which allows it to learn more complex data representations. In contrast, the
network’s width refers to the number of neurons in each layer. A larger network has more
neurons, allowing it to learn more detailed data information [46].
As a result, the depth and the width of the network are determined by the available
computational resources, where the model will be deployed. The four models (YOLOX-Ray-
s, YOLOX-Ray-m, YOLOX-Ray-l and YOLOX-Ray-x) were created by changing the network
depth and width values in order to provide a set of models with different computational
requirements and expected performance. The lightest and fastest model (YOLOX-Ray-s)
has the lowest expected mAP values. The largest model, on the other hand (YOLOX-Ray-x),
is the heaviest and slowest, but has the best expected performance in terms of mAP.
Table 4 presents the network depth and width values for each model size.

Table 4. Depth and width values for each YOLOX-Ray model size.

Small (s) Medium (m) Large (l) Extra-Large (x)

Depth 0.33 0.67 1.0 1.33
Width 0.50 0.75 1.0 1.25

The depth and width values presented in Table 4 were derived from the model scaling
techniques proposed in YOLOv5 by Ultralytics [19] and subsequently adopted in YOLOX
by its authors [23]. In this work, the same logic for model scaling was applied to define
Sensors 2023, 23, 4681 14 of 26

the values for depth and width, considering different model sizes of the YOLOX-Ray
architecture.

4.4. Performance Metrics

For the evaluation metrics, the IoU score is used as a threshold for determining
whether a prediction is considered a True Positive (TP), True Negative (TN), False Positive
(FP) or False Negative (FN). For example, if the IoU score between a predicted bounding
box and the corresponding GT bounding box is greater than a certain threshold (e.g., 0.5),
the prediction is considered a TP. On the other hand, if the IoU score is below the threshold,
the prediction is considered an FP [47].
In this paper, the YOLOX-Ray models in terms of Precision (P), Recall (R), mAP over
an IoU score of 0.5 (mAP50 ), mAP on an IoU threshold of 0.5 to 0.95 (mAP50:95 )

TP
Precision = (15)
TP + FP
TP
Recall = (16)
TP + FN
where FN is the number of false negative detections, TP is the number of correctly predicted
positive instances and FP is the number of false positive predictions.
For calculating AP [47], Equation (17) is used,
Z 1
AP = p(r )dr (17)
0
Equation (18) is used for calculating mAP scores.
N
1
mAP =
N ∑ APi (18)
i =1

where N is the number of classes in the target dataset and APi is the average precision for
class i.
The mAP metric is widely used as a primary evaluation measure in object detection.
It provides an overall evaluation of the performance of an object detection algorithm by
incorporating precision and recall information. The mAP metric is used to compare the
performance of various algorithms on well-known benchmark datasets such as COCO and
PASCAL VOC. This metric has been widely adopted as a standard for comparing different
object detection algorithms and it has been featured in numerous research publications [47].

4.5. Experimental Results

The YOLOX-Ray architecture’s experimental tests were carried out on three distinct
case studies, as previously outlined in the present section. Consequently, the perfor-
mance of the YOLOX-Ray architecture was evaluated across four different model sizes.
These various model sizes were analyzed to find the optimal trade-off between performance
and computational efficiency.
Conducting experimental results for different model sizes in different use cases (Case
studies A, B and C) is essential for assessing the YOLOX-Ray architecture in a variety of
real-world situations since different use cases present unique challenges and requirements.
The YOLOX-Ray architecture must be resilient, robust and effective in detecting anomalies
of varying sizes and shapes in different environments, which can only be achieved through
testing on a range of use cases.
Furthermore, this section includes a comparison of image predictions (object detection)
for each case study and every YOLOX-Ray model size. These images display the YOLOX-
Ray architecture’s capacity to detect and to localize objects within images. The object
detection scores are presented as bounding boxes surrounding each detected object, with
the scores indicating the confidence level that the object belongs to the identified class.
Sensors 2023, 23, 4681 15 of 26

The images illustrate the YOLOX-Ray architecture’s performance in various industrial

inspection use cases and with different model sizes.
The images were selected from the test subset of each case study and they represent
only a single example prediction. Other predictions were made, but only the presented ones
were chosen to emphasize certain strengths, limitations and differences of the YOLOX-Ray
models.
Figures 5–7 have four images, (a), (b), (c) and (d), which are the same image but with
different detection scores, each for a different model. Image (a) illustrates the detection
scores for the smallest model, YOLOX-Ray-s, while image (b) depicts the medium-sized
model, YOLOX-Ray-m. Image (c) depicts the large model, YOLOX-Ray-l and image (d)
illustrates the detection scores for the extra-large model, YOLOX-Ray-x.The evaluation
results are presented in Tables 5–7, each showing the performance of the YOLOX-Ray
models in terms of P, R, mAP over an IoU score of 0.5 (mAP50 ), mAP on an IoU threshold
of 0.5 to 0.95 (mAP50:95 ), inference times in ms (In f .) and the number of parameters in
millions (Params).

4.6. Case Study A: Experimental Results and Predictions

Table 5 demonstrates the evaluation metrics and their values for Case Study A.

Table 5. Performance evaluation of the models on Case Study A.

Model P R mAP50 mAP50:95 Inf. (ms) Params (106 )

YOLOX-Ray-s 0.73 0.917 0.877 0.422 11.95 8.94
YOLOX-Ray-m 0.829 0.915 0.872 0.426 19.55 25.28
YOLOX-Ray-l 0.806 0.916 0.89 0.427 29.22 54.15
YOLOX-Ray-x 0.733 0.879 0.845 0.376 46.56 99.0

Examining Table 5 and beginning with the P metric, the YOLOX-Ray-m and YOLOX-
Ray-l models achieved higher values (0.829 and 0.806, respectively) compared to the small
model (0.73). This indicates that the medium and large models are more accurate in
detecting hotspots. Curiously, the extra-large model had one of the lowest p values (0.733).
In terms of R, all models achieved high values, with YOLOX-Ray-s reaching the
highest at 0.917 and YOLOX-Ray-x obtaining the lowest at 0.879. This suggests that the
models were successful in identifying most anomalies present in the images, regardless of
their size.
In terms of mAP50 , YOLOX-Ray-l performed the best with a value of 0.89, followed
by YOLOX-Ray-s with 0.877. YOLOX-Ray-m and YOLOX-Ray-x achieved similar scores
(0.872 and 0.845, respectively), with the extra-large model having the lowest score. This
implies that larger models may not be optimal for this specific use case.
Regarding mAP50:95 , YOLOX-Ray-l achieved the highest score of 0.427, closely fol-
lowed by YOLOX-Ray-s and YOLOX-Ray-m with 0.422 and 0.426, respectively. YOLOX-
Ray-x obtained the lowest score of 0.376. This metric indicates that YOLOX-Ray-l and
YOLOX-Ray-s are the most accurate models in detecting hotspots with a high IoU score.
Inference time is a crucial factor in real-time object detection applications. In this
instance, YOLOX-Ray-s had the fastest inference time at 11.95 ms, followed by YOLOX-
Ray-m (19.55 ms), YOLOX-Ray-l (29.22 ms) and YOLOX-Ray-x (46.56 ms). As expected,
this suggests that smaller models are more efficient regarding inference time, making them
better suited for real-time object detection.
Finally, the number of parameters for each model varied significantly. YOLOX-Ray-s
had the fewest parameters with 8.94 million, followed by YOLOX-Ray-m with 25.28 million,
YOLOX-Ray-l with 54.15 million and YOLOX-Ray-x with 99 million. This indicates that
smaller models are more lightweight and may be more appropriate for resource-limited
environments.
As expected, the inference time and number of parameters for each model also in-
creased as the model’s size grew. Overall, the YOLOX-Ray architecture demonstrated
Sensors 2023, 23, 4681 16 of 26

solid performance in this case study, which allows for potential further improvement if the
model size is not a concern.
Figure 5 illustrates the instance predictions of different models for Case Study A.

(a) (b)

(c) (d)

Figure 5. Image predictions for Case Study A: (a) Prediction on YOLOX-Ray-s; (b) Prediction on
YOLOX-Ray-m; (c) Prediction on YOLOX-Ray-l; (d) Prediction on YOLOX-Ray-x.

By analyzing Figure 5, it can be concluded that this image contains only small instances
of the class ’Fault’, allowing the evaluation of the YOLOX-Ray architecture’s capacity to
detect small objects.
It is noticeable that the YOLOX-Ray-s model had the lowest prediction scores for
all detections. In contrast, in line with the results obtained and presented in Table 5,
the medium and large models achieved the best detection scores, with YOLOX-Ray-m
achieving the highest prediction scores. Interestingly, the YOLOX-Ray-x model did not
perform well in this specific example, illustrating that even models designed to excel can
underperform compared to lighter models in certain situations.
In summary, this example demonstrates the effectiveness of the YOLOX-Ray architec-
ture in detecting small objects and emphasizes the significance of choosing the suitable
model size based on the task requirements and dataset characteristics.

4.7. Case Study B: Experimental Results and Predictions

Table 6 demonstrates the evaluation metrics and their values for Case Study B.

Table 6. Performance evaluation of the models on Case Study B.

Params
Model P R mAP50 mAP50:95 Inf. (ms)
(106 )
YOLOX-Ray-s 0.984 0.987 0.996 0.66 9.62 8.94
YOLOX-Ray-m 0.972 0.975 0.994 0.661 17.09 25.28
YOLOX-Ray-l 0.962 0.979 0.994 0.658 25.96 54.15
YOLOX-Ray-x 0.972 0.971 0.977 0.625 42.53 99.0
Sensors 2023, 23, 4681 17 of 26

Table 6 shows the YOLOX-Ray-s model surpassed all other models in all metrics,
except for a slightly lower value of mAP50:95 when compared to the medium version
(0.66 vs. 0.661). This indicates that the smaller model is adequate for achieving high
performance in the crack detection task and implies that a smaller model can be a more
effective solution in terms of both inference time and model complexity.
In terms of P, YOLOX-Ray-s also outperformed the other models, achieving a value
of 0.984. The medium and large versions had slightly lower values of 0.972 and 0.962,
respectively, while the extra-large version achieved a p value of 0.972.
All models secured high R values, ranging between 0.971 and 0.987. The small model
achieved the highest value of 0.987, followed by the large version with a value of 0.979.
Regarding mAP50 and mAP50:95 , all models secured high values, spanning from 0.977
to 0.996 for mAP50 and 0.625 to 0.661 for mAP50:95 . The YOLOX-Ray-s model achieved
the highest values for mAP50 , while the medium version achieved the highest value for
mAP50:95 .
In terms of inference times, it is worth noting that YOLOX-Ray-s reached the lowest
inference time, with a value of 9.62 ms, followed by YOLOX-Ray-m and YOLOX-Ray-l at
17.09 and 25.96 ms, respectively. As expected, the YOLOX-Ray-x model had the highest
inference time, with a value of 42.53 ms.
Finally, it is important to highlight that the number of parameters remained the same
across all models, since they were trained using the same configuration. The only difference
was the dataset that was used for training and evaluation.
To conclude, the YOLOX-Ray-s model demonstrated the best overall performance in
the crack detection task, outperforming the larger and more complex models in terms of
both mAP50 and inference times. These results suggest that smaller models can be a feasible
solution for this industrial inspection task, particularly when efficiency is the key.
Figure 6 illustrates the instance predictions of different models for Case Study B.

(a) (b)

(c) (d)

Figure 6. Image predictions for Case Study B: (a) Prediction on YOLOX-Ray-s; (b) Prediction on
YOLOX-Ray-m; (c) Prediction on YOLOX-Ray-l; (d) Prediction on YOLOX-Ray-x.
Sensors 2023, 23, 4681 18 of 26

By analyzing Figure 6, it is possible to conclude that the YOLOX-Ray-s model produced

a false positive detection, which is a significant observation, suggesting that smaller models
might be more susceptible to false positives. The medium model (YOLOX-Ray-m) achieved
the highest prediction score of 80%, followed closely by the YOLOX-Ray-x at 74.4%.
Additionally, the YOLOX-Ray-m achieved a more accurate bounding box regression
aligned with the GT box compared to other models, implying that the YOLOX-Ray-m
model is better at fitting the ’crack’ instance.
These results show that, while all models performed almost identically in terms of
evaluation metrics (as shown in Table 6), false positives can still occur in weaker models.
This emphasizes the importance of selecting a suitable model for each use case scenario, as
well as the need to investigate additional methods for reducing false positives in smaller
models.

4.8. Case Study C: Experimental Results and Predictions

Table 7 demonstrates the evaluation metrics and their values for Case Study C.

Table 7. Performance evaluation of the models on Case Study C.

Model P R mAP50 mAP50:95 Inf. (ms) Params (106 )

YOLOX-Ray-s 0.762 0.866 0.859 0.484 18.04 8.94
YOLOX-Ray-m 0.829 0.878 0.871 0.499 26.60 25.28
YOLOX-Ray-l 0.792 0.883 0.873 0.505 37.83 54.15
YOLOX-Ray-x 0.832 0.876 0.877 0.518 58.12 99.0

Table 7 displays the experimental outcomes of the YOLOX-Ray models trained and
evaluated on Case Study C, which is more challenging than the other two case studies due
to the presence of three classes: ’slippage’, ’corrosion’ and ’crack’.
First, considering the P metric, the YOLOX-Ray-x model achieved the highest value of
0.832, indicating that it made fewer false positive predictions compared to other models.
The YOLOX-Ray-m and YOLOX-Ray-l models also demonstrated high p values, at 0.829
and 0.792, respectively. However, the YOLOX-Ray-s model had the lowest p at 0.762,
signifying a higher rate of false positives.
Next, examining R, which evaluates the model’s ability to accurately identify positive
instances, the YOLOX-Ray-l model obtained the highest value of 0.883. The YOLOX-Ray-
m and YOLOX-Ray-x models also performed well in R, with values of 0.878 and 0.876,
respectively. The YOLOX-Ray-s model had the lowest R at 0.866.
Regarding mAP50 , the YOLOX-Ray-x model reached the highest value of 0.877.
The YOLOX-Ray-m and YOLOX-Ray-l models also posted high mAP50 values, at 0.871
and 0.873, respectively. However, the YOLOX-Ray-s model had the lowest mAP50 at 0.859,
indicating a lower average P across all thresholds.
Lastly, for mAP50:95 , representing the mean average P with a threshold range of
0.50 to 0.95, the YOLOX-Ray-x model achieved the highest value of 0.518. The YOLOX-
Ray-l model also had a relatively high mAP50:95 value of 0.505. The YOLOX-Ray-m and
YOLOX-Ray-s models recorded values of 0.499 and 0.484, respectively.
In terms of inference times, the YOLOX-Ray-s model had the shortest time at 18.04 ms,
while the YOLOX-Ray-x model had the longest time at 58.12 ms. This is expected, as larger
models require more computation time.
Figure 7 illustrates the instance predictions of different models for Case Study C.
Sensors 2023, 23, 4681 19 of 26

(a) (b)

(c) (d)

Figure 7. Image predictions for Case Study C: (a) Prediction on YOLOX-Ray-s; (b) Prediction on
YOLOX-Ray-m; (c) Prediction on YOLOX-Ray-l; (d) Prediction on YOLOX-Ray-x.

Examining Figure 7, it is evident that this figure includes multiple instances of the
‘crack’ and ‘corrosion’ classes, which the YOLOX-Ray models were expected to accurately
detect.
By analyzing image (a), it is possible to conclude that the YOLOX-Ray-s model missed
three ‘crack’ instances in the image, indicating room for enhancement in its detection
abilities.
The YOLOX-Ray-m model’s performance was slightly inferior to that of the YOLOX-
Ray-s model, as it misidentified a ‘crack’ instance as a ‘corrosion’ instance.
The YOLOX-Ray-l model achieved better prediction scores than the YOLOX-Ray-s
and YOLOX-Ray-m models but still failed to identify three ‘crack’ instances in the image.
On the other hand, the YOLOX-Ray-x model, despite having lower prediction scores,
successfully detected all instances in the image, making it the only model to achieve 100%
object detection for this specific image.
This example underlines the variations in detection capabilities among the YOLOX-
Ray models and the trade-offs between prediction scores and detection performance.
Although the YOLOX-Ray-x model achieved perfect object detection, its prediction scores
were lower than those of the YOLOX-Ray-l model.
Moreover, the YOLOX-Ray-s and YOLOX-Ray-m models had lower prediction scores
than the YOLOX-Ray-l model but missed certain instances, signifying the necessity for
model enhancements.
Overall, this example shows the importance of striking a balance between prediction
scores and detection performance in object detection models and the need for ongoing
research and development to improve model capabilities.

4.9. Ablation Study

Ablation studies play a crucial role in DL experiments as they help to determine the
contributions of specific techniques, features or components added to a DL base architecture
to enhance its overall performance [48].
Sensors 2023, 23, 4681 20 of 26

The objective of this study is to compare the YOLOX-Ray results across all case studies
when the SimAM attention mechanism is added to the YOLOX base architecture, the
Alpha-IoU loss function is implemented and finally the YOLOX-Ray architecture, which is
a combination of SimAM and Alpha-IoU.
Experiments were conducted using the smallest model (YOLOX-s) in all case studies,
with the evaluated metrics being P, R, mAP50 , mAP50:95 , inference time in ms (In f .) and
Frames Per Second (FPS). The ablation study results can be visualized in Tables 8–10.
Incorporating the additional components into our model has not led to a change in the
number of parameters. Consequently, the computational cost remains relatively unaffected
by these enhancements. Therefore, since we are only using the smallest model, the number
of parameters is fixed in 8.94 million.
Table 8 represents the evaluation metrics and their values for Case Study A.

Table 8. Ablation study results for Case Study A.

Configuration P R mAP50 mAP50:95 Inf. (ms) FPS

YOLOX 0.77 0.91 0.857 0.40 11.78 84.89
Attention Mechanisms
YOLOX + SENet 0.397 0.891 0.827 0.332 12.14 82.37
YOLOX + CBAM 0.431 0.872 0.797 0.315 12.46 80.26
YOLOX + CA 0.468 0.888 0.828 0.324 12.05 82.99
YOLOX + SimAM 0.359 0.916 0.861 0.371 12.32 81.17
Loss Functions
YOLOX + CIoU 0.601 0.913 0.871 0.378 11.89 84.10
YOLOX + DIoU 0.551 0.9 0.84 0.34 12.44 80.38
YOLOX + GIoU 0.466 0.885 0.823 0.315 11.97 83.54
YOLOX + Alpha-IoU 0.464 0.915 0.866 0.378 12.18 82.10
Proposed Method
YOLOX-Ray 0.73 0.917 0.877 0.422 11.95 83.68

By analyzing Table 8, it is possible to observe that the YOLOX-Ray configuration,

which integrates both the SimAM attention mechanism and Alpha-IoU loss function, out-
performs all other configurations in all metrics, except for P and inference time.
Moreover, it achieved a high R value of 0.917, the highest mAP50 value of 0.877 and
the highest mAP50:95 value of 0.422 when compared to other configurations. Nevertheless,
its p value of 0.73 was slightly lower than the YOLOX configuration. In terms of speed, the
YOLOX-Ray configuration boasted a relatively high FPS value of 83.68 and a low inference
time of 11.95.
The YOLOX configuration secured the highest p value of 0.77, but possessed a mAP50
value of 0.857 and an mAP50:95 value of 0.40. It also demonstrated the highest FPS value of
84.89 and the lowest inference time of 11.78 ms, indicating rapid processing speed in this
case study.
Regarding the other configurations with alternative attention mechanisms, YOLOX
+ SENet achieved a p value of 0.397, an R value of 0.891 and a mAP50 value of 0.827.
YOLOX + CBAM reported a p value of 0.431, an R value of 0.872 and a mAP50 value
of 0.797. Lastly, YOLOX + CA obtained a p value of 0.468, an R value of 0.888 and a
mAP50 value of 0.828. Among these, the YOLOX + CA configuration demonstrated
the best performance in terms of P and R, while the YOLOX + SimAM configuration
achieved the highest mAP50 value in all attention mechanisms.
For the configurations with alternative loss functions, YOLOX + CIoU achieved
a p value of 0.601, an R value of 0.913 and a mAP50 value of 0.871. YOLOX + DIoU
obtained a p value of 0.551, an R value of 0.9 and a mAP50 value of 0.84. Lastly, YOLOX
Sensors 2023, 23, 4681 21 of 26

+ GIoU reported a p value of 0.466, an R value of 0.885 and a mAP50 value of 0.823.
Among these, the YOLOX + CIoU configuration demonstrated the best performance
in terms of P, R and mAP50 values.
In conclusion, the ablation study results for Case Study A reveal that the YOLOX-Ray
configuration delivered the best overall performance among all configurations, despite
having a slightly lower p value than the YOLOX configuration. Such observations al-
low us to state that the combination of the SimAM attention mechanism and the Alpha-
IoU loss function can effectively enhance the YOLOX-Ray architecture’s performance.
However, the specific performance of each configuration depends on the task and dataset
characteristics and the balance between speed and mAP must be considered when choosing
the appropriate configuration.
Table 9 represents the evaluation metrics and their values for Case Study B.

Table 9. Ablation study results for Case Study B.

Configuration P R mAP50 mAP50:95 Inf. (ms) FPS

YOLOX 0.67 0.931 0.897 0.33 9.59 104.28
Attention Mechanisms
YOLOX + SENet 0.699 0.76 0.821 0.357 10.45 95.69
YOLOX + CBAM 0.719 0.88 0.845 0.361 9.98 100.2
YOLOX + CA 0.722 0.961 0.963 0.555 9.7 103.09
YOLOX + SimAM 0.97 0.986 0.99 0.634 9.71 102.99
Loss Functions
YOLOX + CIoU 0.933 0.98 0.989 0.61 9.72 102.88
YOLOX + DIoU 0.913 0.966 0.951 0.581 9.81 101.94
YOLOX + GIoU 0.912 0.964 0.911 0.563 9.86 101.41
YOLOX + Alpha-IoU 0.957 0.972 0.976 0.569 9.91 100.91
Proposed Method
YOLOX-Ray 0.984 0.987 0.996 0.66 9.62 103.95

Table 9 demonstrates that the YOLOX-Ray configuration achieved the highest values
across all metrics, with the exception of inference time. The configuration obtained the
highest mAP50 value of 0.996, the highest mAP50:95 value of 0.66, the highest p value of
0.984 and the highest R value of 0.987. However, it experienced a slightly higher inference
time of 9.62 ms and a lower FPS value of 103.95 compared to the YOLOX base configuration.
Regarding the attention mechanisms, the YOLOX + SENet configuration achieved
a p value of 0.699, an R value of 0.76, an mAP50 value of 0.821 and an mAP50:95 value of
0.357. The YOLOX + CBAM configuration reached a p value of 0.719, an R value of 0.88,
an mAP50 value of 0.845 and an mAP50:95 value of 0.361. The YOLOX + CA configuration
obtained a p value of 0.722, an R value of 0.961, an mAP50 value of 0.963 and an mAP50:95
value of 0.555. Among these, the YOLOX + SimAM configuration demonstrated the best
performance in terms of P, R and mAP values.
For the configurations with alternative loss functions, the YOLOX + CIoU con-
figuration achieved a p value of 0.933, an R value of 0.98, an mAP50 value of 0.989 and
an mAP50:95 value of 0.61. The YOLOX + DIoU configuration obtained a p value of
0.913, an R value of 0.966, an mAP50 value of 0.951 and an mAP50:95 value of 0.581. The
YOLOX + GIoU configuration reported a p value of 0.912, an R value of 0.964, an mAP50
value of 0.911 and an mAP50:95 value of 0.563. Among these, the YOLOX + Alpha-IoU
configuration demonstrated the best performance in terms of P, R and mAP values.
In summary, the YOLOX-Ray configuration is the best choice for object detection
in the crack detection case study, as it achieved the highest values in almost all metrics
except inference time. The YOLOX base configuration is not recommended due to its poor
Sensors 2023, 23, 4681 22 of 26

performance in most metrics. While the addition of SimAM or Alpha-IoU improved certain
metrics individually, the combination of both led to a better performance. It is important
to note that lower inference times are preferred in real-time applications and higher FPS
values signify the model’s capacity to process images more rapidly. Consequently, the
YOLOX-Ray configuration demonstrated superior performance in terms of mAP while
maintaining exceptional performance in terms of speed.
Table 10 represents the evaluation metrics and their values for Case Study C.

Table 10. Ablation study results for Case Study C.

Configuration P R mAP50 mAP50:95 Inf. (ms) FPS

YOLOX 0.29 0.821 0.768 0.389 17.47 57.24
Attention Mechanisms
YOLOX + SENet 0.585 0.87 0.856 0.445 17.90 55.87
YOLOX + CBAM 0.577 0.861 0.849 0.422 18.50 54.05
YOLOX + CA 0.521 0.866 0.858 0.439 18.58 53.81
YOLOX + SimAM 0.277 0.871 0.84 0.45 17.47 57.24
Loss Functions
YOLOX + CIoU 0.657 0.857 0.846 0.435 18.33 54.56
YOLOX + DIoU 0.649 0.857 0.846 0.437 18.16 55.07
YOLOX + GIoU 0.604 0.846 0.825 0.427 17.64 56.69
YOLOX + Alpha-IoU 0.304 0.841 0.806 0.402 18.20 54.95
Proposed Method
YOLOX-Ray 0.762 0.866 0.859 0.484 18.04 55.43

By analyzing Table 10, it becomes clear that the YOLOX-Ray configuration surpassed
all other configurations regarding the two most challenging metrics, mAP50 and mAP50:95 ,
obtaining values of 0.859 and 0.484, respectively. Nevertheless, it had a slightly lower R
value of 0.866 compared to the YOLOX + SimAM configuration, suggesting that it failed
to detect some true positive objects. This configuration also had a relatively low inference
time of 18.04 ms and a relatively high FPS value of 55.43, making it slightly slower than the
quickest configurations (YOLOX and YOLOX + SimAM).
The YOLOX base configuration exhibited the weakest performance in nearly all met-
rics, with a p value of 0.29, an R value of 0.821, an mAP50 value of 0.768, an mAP50:95 value
of 0.389. However, it had the lowest inference time (17.47 ms) and consequently the highest
FPS value (57.24). Curiously, despite being the simplest architecture, it achieved the same
inference time as the YOLOX + SimAM configuration.
For the attention mechanisms, both the YOLOX + SimAM and YOLOX + CA
configurations achieved lower values of mAP50 and mAP50:95 in comparison to the
YOLOX-Ray configuration. Specifically, the YOLOX + SimAM configuration reached a
slightly lower mAP50 value of 0.84 and a lower mAP50:95 value of 0.45 when compared to
the YOLOX-Ray configuration, while the YOLOX + CA configuration secured a slightly
higher mAP50 value of 0.858 and a lower mAP50:95 value of 0.439 when compared to the
YOLOX-Ray configuration. Both configurations had high R values relative to YOLOX-
Ray, with YOLOX + SimAM exhibiting the highest R value of 0.871 and YOLOX + CA
displaying an R value of 0.866. Concerning inference time and FPS, YOLOX + SimAM
had also the lowest inference time of 17.47 ms and, consequently, the highest FPS value
of 57.24 among all configurations, while YOLOX + CA had a slightly longer inference
time of 18.58 ms and a lower FPS value of 53.81.
Regarding the loss functions, the YOLOX + Alpha-IoU configuration achieved
lower values of mAP50 and mAP50:95 in comparison to the YOLOX-Ray configuration,
securing an even lower mAP50 value of 0.806 and a lower mAP50:95 value of 0.402.
Sensors 2023, 23, 4681 23 of 26

The YOLOX + Alpha-IoU configuration had a high R value relative to YOLOX-Ray,

displaying an R value of 0.841. Concerning inference time and FPS, YOLOX + Alpha-
IoU had a slightly longer inference time of 18.20 ms and a lower FPS value of 54.95.
In summary, the YOLOX-Ray configuration delivered the best performance in terms of
the most crucial metrics (mAP50 and mAP50:95 ), despite having a lower R value and slower
inference times compared to some other configurations. The results also imply that, in
this case study, faster inference times and higher FPS values are preferable but should not
undermine model performance. The attention mechanisms and loss functions individually
showed improvements over the base YOLOX configuration, but the combination of these
techniques in the YOLOX-Ray configuration led to the most significant performance gains.

5. Conclusions
The experimental results presented in this study demonstrate the YOLOX-Ray deep
learning architecture’s effectiveness in the context of industrial inspection tasks. We demon-
strated the YOLOX-Ray model’s versatility in a wide range of scenarios by evaluating its
performance across three distinct case studies of variable complexity.
The ablation study’s findings revealed that incorporating SimAM and Alpha-IoU
individually improved the performance of the YOLOX base architecture. However, when
combining both components (YOLOX-Ray) the best overall results can be achieved. These
findings highlight the importance of selecting and fine-tuning the various components of
deep learning architectures for specific use cases. They can also serve as a foundation for
future industrial inspection research and development.
The YOLOX-Ray architecture improves object detection in real-world industrial inspec-
tion scenarios by incorporating both the SimAM attention mechanism and the Alpha-IoU
loss function. By incorporating the SimAM module, the architecture can effectively refine
features in both the channel and spatial domains, resulting in a more accurate and robust ob-
ject detection performance. SimAM’s lightweight and efficient design allows for increased
computational efficiency while maintaining detection quality. The Alpha-IoU loss function,
on the other hand, addresses the challenges of detecting small objects and scale variations,
making it suitable for difficult industrial inspections and anomaly detection tasks.
Due to the combination of the SimAM attention mechanism and the Alpha-IoU loss
function, the YOLOX-Ray architecture performs better, in particular, on complex and chal-
lenging environments, ensuring more precise detections and improved overall performance.
The YOLOX-Ray architecture is an optimal choice for tackling real-world object detection
tasks and applications due to its powerful fusion of attention mechanism and loss function.

Author Contributions: Conceptualization, A.R.; methodology, A.R.; software, A.R.; validation, A.R.;
formal analysis, J.P.P.; investigation, A.R. and J.P.P.; resources, P.S. and O.P.; data curation, A.R.;
writing—original draft preparation, A.R.; writing—review and editing, J.P.P., P.S. and O.P.; visualiza-
tion, A.R.; supervision, P.S. and O.P.; project administration, P.S. and O.P.; funding acquisition, P.S.
and O.P. All authors have read and agreed to the published version of the manuscript.
Funding: This work was supported by Instituto de Telecomunicações Lisbon, Portugal and FCT/MCTES
through national funds and when applicable co-funded EU funds under the project UIDB/50008/2020.
Additionally the work was supported by Iscte—Instituto Universitário de Lisboa, under the grant
ISTA-BM-PDCTI-2017.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: The YOLOX-Ray source code, as well as the model’s weights for
pre-training other custom datasets can be found on this https://github.com/Sphincz/YOLOX-Ray
(GitHub repository). The datasets used in the case studies can be found in the https://www.rf100.org/
(Roboflow-100) official website.
Sensors 2023, 23, 4681 24 of 26

Acknowledgments: We would like to thank Iscte—Instituto Universitário de Lisboa and

IT—Telecommunications Institute for providing the needed hardware resources to perform
this research.
Conflicts of Interest: The authors declare no conflict of interest.

Abbreviations
The following abbreviations are used in this manuscript:

AI Artificial Intelligence
AP Average Precision
CA Coordinate-Attention
CAM Channel Attention Mechanism
CBAM Convolutional Block Attention Module
CCA Coordinate-Channel Attention
CE Cross-Entropy
CIoU Complete Intersection-over-Union
CNN Convolutional Neural Network
COCO Common Objects in Context
CPU Central Process Unit
CSA Coordinate-Spatial Attention
CSP Cross Stage Partial
CV Computer Vision
DIoU Distance Intersection-over-Union
DL Deep Learning
DNN Deep Neural Network
DOTA Dataset for Object Detection in Aerial images
FPN Feature Pyramid Network
FPS Frames Per Second
GIoU Generalized Intersection-over-Union
GPU Graphics Processing Unit
GT Ground-Truth
HSV Hue, Saturation and Value
IoU Intersection-over-Union
mAP Mean Average Precision
ML Machine Learning
OTA Optimal Transport Assignment
PAN Path Aggregation Network
RPN Region Proposal Network
SAM Spatial Attention Mechanism
SE Squeeze-and-Excitation
SENet Squeeze-and-Excitation Network
SimAM Simple Parameter-free Attention Module
SimOTA Simplified Optimal Transport Assignment
SoTA State-of-the-Art
VOC Visual Object Classes
YOLO You Only Look Once

References
1. Kumar, A. Computer-Vision-Based Fabric Defect Detection: A Survey. IEEE Trans. Ind. Electron. 2008, 55, 348–363. [CrossRef]
2. Weimer, D.; Scholz-Reiter, B.; Shpitalni, M. Design of deep convolutional neural network architectures for automated feature
extraction in industrial inspection. CIRP Ann. 2016, 65, 417–420. [CrossRef]
3. Bedi, P.; Goyal, S.B.; Rajawat, A.S.; Bhaladhare, P.; Aggarwal, A.; Prasad, A. Feature Correlated Auto Encoder Method for
Industrial 4.0 Process Inspection Using Computer Vision and Machine Learning. Procedia Comput. Sci. 2023, 218, 788–798.
[CrossRef]
4. Voulodimos, A.; Doulamis, N.; Doulamis, A.; Protopapadakis, E. Deep Learning for Computer Vision: A Brief Review. Comput.
Intell. Neurosci. 2018, 2018, 7068349.
Sensors 2023, 23, 4681 25 of 26

5. Redmon, J.; Divvala, S.; Girshick, R.; Farhadi, A. You Only Look Once: Unified, Real-Time Object Detection. In Proceedings of the
2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 27–30 June 2016; pp. 779–788.
[CrossRef]
6. Liu, W.; Anguelov, D.; Erhan, D.; Szegedy, C.; Reed, S.; Fu, C.Y.; Berg, A.C. SSD: Single Shot MultiBox Detector. CoRR 2016, 21–37.
[CrossRef]
7. Lin, T.Y.; Goyal, P.; Girshick, R.; He, K.; Dollar, P. Focal Loss for Dense Object Detection. IEEE Trans. Pattern Anal. Mach. Intell.
2020, 42, 318–327. [CrossRef] [PubMed]
8. Ren, S.; He, K.; Girshick, R.; Sun, J. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. IEEE
Trans. Pattern Anal. Mach. Intell. 2017, 39, 1137–1149. [CrossRef]
9. He, K.; Gkioxari, G.; Dollar, P.; Girshick, R. Mask R-CNN. In Proceedings of the 2017 IEEE International Conference on Computer
Vision (ICCV), Venice, Italy, 22–29 October 2017; pp. 2980–2988. [CrossRef]
10. Cai, Z.; Vasconcelos, N. Cascade R-CNN: Delving Into High Quality Object Detection. In Proceedings of the 2018 IEEE/CVF
Conference on Computer Vision and Pattern Recognition, Salt Lake City, UT, USA, 18–23 June 2018; pp. 6154–6162. [CrossRef]
11. Du, J. Understanding of Object Detection Based on CNN Family and YOLO. J. Phys. Conf. Ser. 2018, 1004, 012029. [CrossRef]
12. Li, Y.; Zeng, J.; Shan, S.; Chen, X. Occlusion Aware Facial Expression Recognition Using CNN with Attention Mechanism. IEEE
Trans. Image Process. 2019, 28, 2439–2450. [CrossRef]
13. Brauwers, G.; Frasincar, F. A General Survey on Attention Mechanisms in Deep Learning. IEEE Trans. Knowl. Data Eng. 2023,
35, 3279–3298. [CrossRef]
14. Mustafaev, B.; Tursunov, A.; Kim, S.; Kim, E. A Novel Method to Inspect 3D Ball Joint Socket Products Using 2D Convolutional
Neural Network with Spatial and Channel Attention. Sensors 2022, 22, 4192. [CrossRef]
15. Yang, L.; Zhang, R.Y.; Li, L.; Xie, X. SimAM: A Simple, Parameter-Free Attention Module for Convolutional Neural Networks.
In Proceedings of the 38th International Conference on Machine Learning, Virtual, 18–24 July 2021; Volume 139, pp. 11863–11874.
16. Cina, M.; Binny, S.; Roji, T.; Cini, J.; Shincy, K.K. Comparison of YOLO Versions for Object Detection from Aerial Images. Int. J.
Eng. Technol. Manag. Sci. 2022, 9, 315–322.
17. Redmon, J.; Farhadi, A. YOLO9000: Better, Faster, Stronger. In Proceedings of the 2017 IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), Honolulu, HI, USA, 21–26 July 2017; pp. 6517–6525. [CrossRef]
18. Redmon, J.; Farhadi, A. YOLOv3: An Incremental Improvement. arXiv 2018, arXiv:1804.02767
19. Ultralytics LLC. YOLOv5. 2020. Available online: https://github.com/ultralytics/yolov5 (accessed on 24 January 2023).
20. Wang, C.Y.; Yeh, I.H.; Liao, H.Y.M. You Only Learn One Representation: Unified Network for Multiple Tasks. arXiv 2021,
arXiv:2105.04206.
21. Ge, Z.; Liu, S.; Wang, F.; Li, Z.; Sun, J. YOLOX: Exceeding YOLO Series in 2021. arXiv 2021, arXiv:2107.08430.
22. Ren, K.; Chen, X.; Wang, Z.; Yan, X.; Zhang, D. Fruit Recognition Based on YOLOX*. Proc. Int. Conf. Artif. Life Robot. 2022,
27, 470–473. [CrossRef]
23. LearnOpenCV. YOLOX Object Detector Paper Explanation and Custom Training. 2022. Available online: https://learnopencv.
com/yolox-object-detector-paper-explanation-and-custom-training/ (accessed on 24 January 2023).
24. Zhang, J.; Huang, B.; Ye, Z.; Kuang, L.D.; Ning, X. Siamese anchor-free object tracking with multiscale spatial attentions. Sci. Rep.
2021, 11, 22908. [CrossRef]
25. Ge, Z.; Liu, S.; Li, Z.; Yoshie, O.; Sun, J. OTA: Optimal Transport Assignment for Object Detection. In Proceedings of the 2021
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Nashville, TN, USA, 20–25 June 2021; pp. 303–312.
[CrossRef]
26. Zhang, H.; Cisse, M.; Dauphin, Y.N.; Lopez-Paz, D. mixup: Beyond Empirical Risk Minimization. arXiv 2017, arXiv:1710.09412.
27. Wei, Z.; Duan, C.; Song, X.; Tian, Y.; Wang, H. AMRNet: Chips Augmentation in Aerial Images Object Detection. arXiv 2020,
arXiv:2009.07168.
28. Zhang, C.; Yang, T.; Yang, J. Image Recognition of Wind Turbine Blade Defects Using Attention-Based MobileNetv1-YOLOv4 and
Transfer Learning. Sensors 2022, 22, 6009. [CrossRef]
29. Hu, J.; Shen, L.; Albanie, S.; Sun, G.; Wu, E. Squeeze-and-Excitation Networks. IEEE Trans. Pattern Anal. Mach. Intell. 2020,
42, 2011–2023. [CrossRef] [PubMed]
30. Woo, S.; Park, J.; Lee, J.Y.; Kweon, I.S. CBAM: Convolutional Block Attention Module. In Lecture Notes in Computer Science
(Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Springer: Cham, Switzerland, 2018;
Volume 11211, pp. 3–19. [CrossRef]
31. Hou, Q.; Zhou, D.; Feng, J. Coordinate Attention for Efficient Mobile Network Design. In Proceedings of the 2021 IEEE/CVF
Conference on Computer Vision and Pattern Recognition (CVPR), Nashville, TN, USA, 20–25 June 2021; pp. 13708–13717.
[CrossRef]
32. Zhai, H.; Cheng, J.; Wang, M. Rethink the IoU-based loss functions for bounding box regression. In Proceedings of the
2020 IEEE 9th Joint International Information Technology and Artificial Intelligence Conference (ITAIC), Chongqing, China,
11–13 December 2020; pp. 1522–1528. [CrossRef]
33. Zheng, Z.; Wang, P.; Liu, W.; Li, J.; Ye, R.; Ren, D. Distance-IoU Loss: Faster and Better Learning for Bounding Box Regression.
Proc. AAAI Conf. Artif. Intell. 2020, 34, 12993–13000. [CrossRef]
Sensors 2023, 23, 4681 26 of 26

34. Li, H.; Zhou, Q.; Mao, Y.; Zhang, B.; Liu, C. Alpha-SGANet: A multi-attention-scale feature pyramid network combined with
lightweight network based on Alpha-IoU loss. PLoS ONE 2022, 17, e0276581. [CrossRef] [PubMed]
35. Rezatofighi, H.; Tsoi, N.; Gwak, J.; Sadeghian, A.; Reid, I.; Savarese, S. Generalized Intersection Over Union: A Metric and a Loss
for Bounding Box Regression. In Proceedings of the 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition
(CVPR), Long Beach, CA, USA, 15–20 June 2019; pp. 658–666. [CrossRef]
36. He, J.; Erfani, S.; Ma, X.; Bailey, J.; Chi, Y.; Hua, X.S. Alpha-IoU: A Family of Power Intersection over Union Losses for Bounding
Box Regression. In Proceedings of the Thirty-Fifth Conference on Neural Information Processing Systems, Online, 6–14 December
2021; Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P.S., Vaughan, J.W., Eds.; Curran Associates, Inc.: Red Hook, NY, USA,
2021; Volume 34, pp. 20230–20242.
37. Xiong, C.; Hu, S.; Fang, Z. Application of improved YOLOV5 in plate defect detection. Int. J. Adv. Manuf. Technol. 2022, 1–13.
38. Liu, L.; Liu, Y.; Yan, J.; Liu, H.; Li, M.; Wang, J.; Zhou, K. Object Detection in Large-Scale Remote Sensing Images With a
Distributed Deep Learning Framework. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2022, 15, 8142–8154. [CrossRef]
39. Bochkovskiy, A.; Wang, C.Y.; Liao, H.Y.M. YOLOv4: Optimal Speed and Accuracy of Object Detection. arXiv 2020,
arXiv:2004.10934.
40. Li, S.; Liu, S.; Cai, Z.; Liu, Y.; Chen, G.; Tu, G. TC-YOLOv5: Rapid detection of floating debris on raspberry Pi 4B. J. Real-Time
Image Process. 2023, 20, 17. [CrossRef]
41. Roboflow. Solar Panels Thermal Dataset. 2022. Available online: https://universe.roboflow.com/neelakshtayal-gmail-com/
thermal-dataset-tfoku (accessed on 10 January 2023).
42. Roboflow. Crack Detection Dataset. 2022. Available online: https://universe.roboflow.com/crack-7rsjb/crack-detection-ol3yi
(accessed on 14 January 2023).
43. Roboflow. Corrosion Detection Dataset. 2022. Available online: https://universe.roboflow.com/roboflow-100/corrosion-bi3q3
(accessed on 14 January 2023).
44. Ciaglia, F.; Zuppichini, F.S.; Guerrie, P.; McQuade, M.; Solawetz, J. Roboflow 100: A Rich, Multi-Domain Object Detection
Benchmark. arXiv 2022, arXiv:2211.13523.
45. Shorten, C.; Khoshgoftaar, T.M. A survey on Image Data Augmentation for Deep Learning. J. Big Data 2019, 6, 60. [CrossRef]
46. Khan, A.; Sohail, A.; Zahoora, U.; Qureshi, A.S. A survey of the recent architectures of deep convolutional neural networks. Artif.
Intell. Rev. 2020, 53, 5455–5516. [CrossRef]
47. Padilla, R.; Passos, W.L.; Dias, T.L.B.; Netto, S.L.; da Silva, E.A.B. A Comparative Analysis of Object Detection Metrics with a
Companion Open-Source Toolkit. Electronics 2021, 10, 279. [CrossRef]
48. Sheikholeslami, S.; Meister, M.; Wang, T.; Payberah, A.H.; Vlassov, V.; Dowling, J. AutoAblation: Automated Parallel Ablation
Studies for Deep Learning. In Proceedings of the 1st Workshop on Machine Learning and Systems, Online, 26 April 2021; ACM:
New York, NY, USA, 2021; pp. 55–61. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.