Fruit Quality Recognition using Deep Learning

Algorithm
Prof.Mrs.Sarika Bobde Prof.Mrs.Pradnya Kulkarni Pranav Khode
Dept.of Computer Engineering, Dept. of Computer Engineering. Dept. of Computer Engineering.
MAEER’s Maharashtra Institute of MAEER’s Maharashtra Institute of MAEER’s Maharashtra Institute of
Technology Technology Technology
(Savitribai Phule Pune University) (Savitribai Phule Pune University) (Savitribai Phule Pune University)
Pune, India Pune, India Pune, India
sarika.bobde@mitwpu.edu.in pradnya.kulkarni@mitpune.edu.in pranavkhode27@gmail.com

Sarthak Jaiswal Omkar Patil Rishabh Jha

Dept. of Computer Engineering. Dept. of Computer Engineering. Dept. of Computer Engineering.
MAEER’s Maharashtra Institute of MAEER’s Maharashtra Institute of MAEER’s Maharashtra Institute of
Technology Technology Technology
(Savitribai Phule Pune University) (Savitribai Phule Pune University) (Savitribai Phule Pune University)
Pune, India Pune, India Pune, India
sjaiswal1204gmail.com omkarpatil32@gmail.com jharishabh1899@gmail.com

Abstract— Fruit classification is essential in various the heart of computer vision, with various techniques and
industrial settings, such as factories, supermarkets, and other methodologies available to conduct the necessary
places. Fruit classification may also be beneficial to persons classification and measurements. Machine vision technology
with unique nutritional needs who utilize it to choose the
is now used to evaluate most commercial fruits, including
proper fruits. Manual sorting was formerly used for fruit
classification is time-consuming and requires continual human
oranges, peaches, and apples, as well as mangoes and
presence. Many fruit classification machine learning bananas. Computer vision techniques are used to classify the
techniques have been proposed in the past. Deep learning may fruit according to the feature based on color, shape, size and
be a powerful engine for generating actionable results in texture.
today's reality because of its detection and classification
abilities. As a result, a convolutional neural network was In today's world, automation is becoming increasingly vital.
employed to construct an effective fruit classification model. It In several industrial domains, computer vision technologies
makes use of the fruits 360 dataset, which contains 131 are rapidly evolving. The quality of fruit can be determined
different fruit and vegetable varieties. In this procedure, we
using image processing techniques. People are health-
used three fruits, divided into three categories: good, raw, and
damaged. The model was made in Keras. It had been trained conscious, and they want only fresh, high-quality fruit. Thus
for 50 epochs and had a 95% accuracy rate. this effort will undoubtedly contribute to the improvement
of fruit quality.
Keywords—classification, machine learning, deep learning, Fruit rotting has substantial economic consequences; it is
sorting, trained, epochs. estimated that roughly a third of the cost of the fruit is spent
on rotting materials. Furthermore, the sale of fruits will be
I. INTRODUCTION impacted because consumers believe that spoiled fruits are
India is primarily a farming country. India produces a wide harmful to their health due to decreased concentrations of
variety of fruits and vegetables. In India, all pre-harvest and amino acids, vitamins, sugar/glucose, and other nutrients, all
post-harvest processes are carried out by hand with the of which inevitably raise public concerns about edibility
assistance of workers. Because manual processes are issues, prompting discussions on how to prevent or slow the
inefficient, automation in the food processing business is decaying process.
required to obtain reliable results. Fruit sorting and grading Fruit freshness grading is significant because of the
are part of the pre-production process. Texture, shape, color, importance of food status in people's lives and contributes to
size, and volume are external quality characteristics, while the economy, yet the manual operation is time-consuming.
test, sweetness, flavor, fragrance, nutrients, and Automation of grading through the use of digital methods is
carbohydrates present in the fruit are internal quality criteria. thought to be the answer to this problem.
A computer vision system is a low-cost system that provides
consistent performance, high speed, and accurate fruit
sorting and grading. Most commercial fruits, such as II. LITERATURE SURVEY
oranges, peaches, and apples, as well as mangoes and An image-based approach for identifying and analyzing fruit
bananas, are now evaluated using machine vision size was proposed by Hongshe Dang, Jinguo Song, and Qin
technology. Guo [1]. On the QT/Embedded platform, the system uses an
A computer vision system is a low-cost system that ARM9 processor as its central processor and image
performs consistently, quickly, and accurately while sorting processing technologies to detect the size of fruits.
and grading fruits. Computer vision is the process of John B. Njoroge et al.[2] created an automated grading
creating clear and meaningful descriptions of actual objects method based on image processing that focuses on the fruit's
from photos. Image processing and picture analysis are at interior and external flaws. Six CCD cameras make up the

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

system. Two cameras are mounted on the fruit's top, two on In [10], the authors offered a system for determining the size
the right, and two on the left. Biological problems are of various fruits and sorting them using fuzzy logic; in this
investigated using X-ray imaging. The fruit's properties are case, the author offered MATLAB for feature extraction and
analyzed using image processing, and the size, color, shape, GUI creation.
and grade are calculated based on the attributes. Advanced [11] suggested Image Processing Techniques for Detection
designs, specific fabrications, and autonomous mechanical of Leaf Disease by Arti N. Rathod (Nov 2013). The
control were used to build the system. investigation looked into various approaches for detecting
J. V. Frances et al. [3] suggested a methodology for leaf disease. The study uses various methodologies to
enhancing the performance of a fruit sorting and grading investigate how image processing might be utilized to detect
machine's load-cell-based weighting subsystem, whether by leaf illness. This study aimed to use image analysis to detect
boosting speed or accuracy, to reach an accuracy of + l the disease-affected region of the cotton leaf sport. The
gramme. study will compare two feature extraction strategies, one of
Wong Bing Yit et al.[4] proposed a new MMS-based system which is threshold clustering and the other is k-means
design for consumer fruit grading based on signal clustering. Using image conversion to HSV color space and
processing. A prototype network architecture, which fuzzy c-means clustering in hue-saturation space, this study
comprised integrating a wireless message system with signal shows how to distinguish distinct pixel classes. During the
processing between mobile consumers, was analyzed, interactive stage, these classes are merged into two final
recommended, and built for development purposes. classes, which are used to choose the disease regions to be
Dah Jye Lee et al. [5] devised a color mapping methodology searched.
for acquiring the color of the fruit. This system has the In [12], Manisha A. Bhange (2015) described an image
advantage of modifying color choices or grading factors processing algorithm for detecting Pomegranate Disease.
according to the application. It is a user-friendly method. The study's author devised a method for identifying
H. Dang et al. [6] developed a technique for detection of pomegranate fruit disease (bacterial blight), which farmers
size of fruit. The process begins with edge detection, can photograph and upload to the system. The equipment
followed by fruit size detection, and finally, fruit grading will tell the farmer whether the fruit has bacterial Bligh. K-
based on size. To obtain a binary image, OSTU is utilized. means clustering is used for image segmentation. The
The 8-connected boundary approach is used to detect edge clusters are sorted into two groups, i.e., infected fruit and
sequences. Symmetry is taken into account when detecting non-diseased fruit image. The main aim of this study was to
diameters. The center coordinate is determined by enhance the efficiency of an automatic fruit disease
symmetry. The diameter of a line is computed using the detection system by employing an intent search technique.
center point and the axis that runs through it. Two edge J. Pujari (2013) employed statistical approaches to detect
points are searched for an accurate result. If the fruit is fruit fungus illness in [13]. Pomegranate, mango, and grapes
turned, the diameter displayed is the same as the diameter were the fruits chosen for research. There are two phases to
that indicated the actual fruit size. Then, based on precise image preparation. The input pre - processed for binarization
size grading, the process is processed. and noise reduction in the first phase. The image is
D. Savakar [7] assessed five distinct types of fruit imaged resize and a bounding box is created in the second phase.
(Apple, Orange, Chickoo, Mango, and Sweet Lemon). A Feature extraction is done using the blockwise feature
total of 5000 images were taken, with 1000 of each fruit extraction methodology. The image is divided into 5*5
species. 18 colors and 27 texture characteristics were blocks. GLCM is used to extract textual features.
extracted to build the technique. Separating the RGB Ab Razak Mansor et al [14] proposed a color sensor modal
components yielded the color characteristics. After that, the that used color as a criteria for mango fruit quality in [14].
RGB picture was transformed to an HSI model and the The fuzzy logic RGB sensor color model built in Matlab
components were separated. Each RGB and HSI was used to develop a unique model for identifying mango
component's mean, variance, and range were computed fruit in this study. The researchers intended to establish a
separately. ripeness index based on RGB color intensity on mango skin,
Based on the obtained attributes, Deepa [8] proposed a develop a equation utilizing color values, and assess the
method for evaluating and classifying defective and non- calibration efficiency by utilizing Fuzzy Logic to categorize
defective fruits. In the database, there were 200 mosambi (i.e., unripe, ripe, and overripe) the mango ripening index.
fruits of different Shape, intensity, and texture were
collected. The database was then classified using PNN, with
the results revealing that shape, intensity, and texture III. SYSTEM ARCHITECTURE
features had achieved detection rate of 100%, 92%, and
96%, respectively.
Mustafa et al. [9] devised a original method for evaluating
fruits classification. Apples, bananas, carrots, mangoes, and
oranges were the focus of this approach. Fruit sample
images were used to extract shape and color data. Apples
and oranges, as well as bananas and carrots, have almost
identical shapes and sizes, therefore morphological
characteristics were employed to identify them. Color Fig. 1 System Architecture
characteristics were employed to minimize misclassification
between apples and oranges, as well as bananas and carrots,
resulting in a 79-90% accuracy.
 IV. FRUIT DATASET V. METHODOLOGY
The fruit image collection was compiled using data from A. Preprocessing
various web sources, including the fruit360 collection. The
dataset has nine classes. Table 1.0 depicts the database .This image has been preprocessed to make it ready for
distribution of the proposed system. further manipulation. The image is subjected to image
processing procedures such as RGB to Gray conversion,
Apple thresholding, and segmentation during preprocessing.
Apple good Apple raw
damaged Filtering is a crucial step in the preprocessing process. The
Training
median filter is used to reduce noise and smooth out the
160 205 132
image in this case.
B. Training and Testing Using CNN
Testing 40 49 32
CNNs are a form of Neural Network that has been
especially good at picture recognition and categorization.
CNN's are a type of feed-forward neural network that has
Lemon many layers. Filters, kernels, or neurons with programmable
Lemon good Lemon raw weights, parameters, and biases make up CNNs. Each filter
damaged
Training takes a set of inputs, performs convolution, and may or may
454 399 277
not include non-linearity. A typical CNN design can be
illustrated in Figure 2.0. CNN's structure is made up of
Testing 114 98 68
convolutional, pooling, Rectified Linear Unit (ReLU), and
Fully Connected Layer.

Mango
Mango good Mango raw
damaged
Training
38 24 21

Testing 14 08 16

TABLE I. Database Distribution

The image samples of the database are as shown in Fig. 2.

Fig. 3. Architecture of CNN

Each block of CNN architecture is explained below

a) Convolutional Layer:
The convolutional layer, the fundamental building
block of a CNN. The main task of this layer is to extract the
features from the input image. It preserve the spatial link
between the image pixels using feature maps or activation
maps. It's written mathematically as.

G [ m, n ]= ( f ∗h ) [ m, n ] =∑ ∑ h [ j, k ] f [m− j , n−k ]
j k

(1)

Fig. 2. Database Samples (a)Apple (damaged) (b)Apple (good) (c)Apple

(raw) (d)Lemon (damaged) (e) Lemon (good) (f) Lemon (raw) (g) Mango
(Damaged) (h) Mango (good) (i) Mango (raw)

Fig. 4. Covolutional Layer

b) ReLu Layer:
 ReLu stands for rectifier-based units in a non-linear Softmax activation function. The output probabilities of the
operation. It's an element-wise operation, which means it's FCL are added together and equal to 1. To ensure this,
done per pixel and replaces any negative feature map values Softmax is utilized as the activation function. The Softmax
with zero. We'll suppose the neuron input is x and the function compresses a vector of arbitrary real-valued scores
rectifier is y to show how the ReLU works. It is defined as into a zero to one value vector that adds up to one.

f (x)=max(0 , x ) (2)

Fig. 6. Fully Connected Layer

Fig. 5. ReLu VI. RESULTS AND ANALYSIS

The proposed fruit grading system's results are assessed both
c) Pooling Layer: qualitatively and quantitatively.
Each activation map's dimensionality is reduced by
the pooling layer while the most critical data is retained. The
input images are divided into a series of non-overlapping A. Qualitative Analysis:
rectangles. Each region is downsampled using a non-linear
approach such as min, max or average. This layer improves The purpose of qualitative analysis is to provide a
generalization, speeds convergence, and resists translation comprehensive and detailed description of the situation.
and distortion. The max-pooling layers are straightforward Rare phenomena are given the same attention as more
and don't require much understanding. common events (or should be given the same amount of
attention). Qualitative analysis allows for nuanced
distinctions because the data does not have to be squeezed
into a finite number of classifications. The analysis is
capable of detecting ambiguity in human language. The
results of fruit grading are shown in Fig 7.

Fig. 4. Max Pooling Layer

d) Flatten Layer:
This layer convert 2D pooled map data into one
column. As the name implies, we'll make a column out of
our pooled feature map, as illustrated in the image below.

Fig. 5. Flattening Layer

e) Fully Connected Layer:

The FCL's objective is to use these features to
classify the input image into various classes using the
training dataset as a reference. FCL, the final pooling layer,
is in charge of feeding data to a classifier that uses the
 (a)

Fig. 7. Output of Fruit Grading System

B. Quantitative Analysis:
To explain what is noticed, quantitative researchers find
traits, count them, and even develop more complex
statistical models. Conclusions can be extended to a larger (b)
population, and direct comparisons between two corpora can
Fig. 8. Quantitative Analysis (a) Accuracy (b) Loss
be made as long as valid sample and significance
methodologies are used. As a result, quantitative analysis
enables us to distinguish between situations that are likely to VII. CONCLUSION
be accurate reflections of a language's or variety's behavior The Convolutional Neural Network-based fruit quality
and those that are merely random instances. The more grading system has been presented in this manner. In our
straightforward process of examining a single fruit range research, we were able to discover two primary application
allows one to obtain a definite picture of the frequency and areas. The first is fruit classification, in which this procedure
rarity of specific events, as well as their relative normality or is directly used to classify fruits by type in market,
abnormality. An accuracy parameter is used to calculate the supermarket, wholesaler, and retailer applications. The
quantitative analysis of the proposed system. The fruit second application is fruit quality control, which detects
quality recognition accuracy is expressed as external damage to fruits or infections. Varied applications,
works, and authors have different designs for CNN-based
No of classes correctly detected techniques.
Accuracy= Furthermore, we cannot declare one CNN architecture to be
Total no of samples
superior to the others. As a result, it is possible to create a
(3) new CNN model using a pre-trained CNN by adjusting
specific layers and parameters and starting from scratch. The
evaluation results demonstrate that CNN-based techniques
obtained excellent results, with up to 95% accuracy.
REFERENCES
[1] Hongshe Dang, Jinguo Song, Qin Guo, “A Fruit Size Detecting and
Grading System Based on Image Processing,” 2010 Second
International Conference on Intelligent Human-Machine Systems and
Cybernetics,pp83-86.
[2] John B. Njoroge. Kazunori Ninomiya. Naoshi Kondo andHideki
Toita, “Automated Fruit Grading System using ImageProcessing,”
The Society of Instrument and Control Engineers(SICE2002), Osaka,
Japan, August 2002, pp 1346-1351.
[3] J. V. Frances, J. Calpe, E. Soria, M. Martinez, A. Rosado,
A.J.Serrano, J. Calleja, M. Diaz, “Application of ARMA modeling to
the improvement of weight estimations in fruit sorting and grading
machinery,” IEEE 2000, pp 3666-3669.
[4] Wong Bing Yit, Nur Badariah Ahmad Mustafa, ZaipatimahAli, Syed
Khaleel Ahmed, Zainul Abidin Md Sharrif, “Design and Development
of a Fully Automated Consumer-based Wireless Communication
System For Fruit Grading”, ISCIT 2009 , pp 364- 369.
[5] D.Lee, J.Archibald and G.Xiong,“ Rapid Color Grading for Fruit
Quality Evaluation Using Direct Color Mapping”, IEEE Transactions
On Automation Science And Engineering, Vol 8, No.2,pp.292-302,
April 2011.
[6] H. Dang, , “A Fruit Size Detecting and Grading System Based on
Image Processing,” 2010 Second International Conference on
Intelligent Human-Machine Systems and Cybernetics,pp83-86.
[7] D. Savakar, “Identification and Classification of Bulk Fruits Images
using Artificial Neural Networks,” in International Journal of
Engineering and Innovative Technology (IJEIT), vol. 1(3), March
2012.
[8] P. Deepa, “A Comparative Analysis of Feature Extraction Methods
for Fruit Grading Classifications,” in International journal of
 emerging technologies in computational and applied sciences [11] Arti N. Rathod (Nov 2013) “Image Processing Techniques for
(IJETCAS) vol.13(138), 2013 Detection of Leaf Disease”.
[9] ] N. B. A. Mustafa, K. Arumugam, S. K. Ahmed,Z. A. M. Sharrif, [12] ] Manisha A. Bhange (2015) “A Review of Image Processing for
“Classification of Fruits using Probabilistic Neural Networks- Pomegranate Disease Detection”.
Improvement using Color Features”, IEEE International Conference [13] Jagadeesh Devdas Pujari (2013) “Grading and Classification of
on TENCON, 2011. Anthracnose Fungal Disease of Fruits based on Statistical Texture
[10] ] Naoshi Kondo, “Fruit Grading Robot”, Proceedings of the2003 Features”.
IEEE/ASME International Conference on AdvancedIntelligent [14] ] Ab Razak Mansor (April 2014) “Fuzzy Ripening Mango Index using
Mechatronics (AIM 2003), pp 1366-1371. [7] R. C. Gonzalez, R. E. RGB color sensor modal”.
Woods," Digital Image Processing", Pearson Education.IIEd.,2002