Visvesvaraya Technological University Belagavi

Project Work First Phase Presentation on

“OPTIMISATION OF NEURAL NETWORK FOR IDENTIFYING SCENE IMAGES”
by
SAMBHRAMA K M: 4JN22EC085
SHIVAPRASAD M : 4JN22EC090
SHREERAM BHAT : 4JN22EC094
THEJUS S K : 4JN22EC114
Under the guidance of
SUMATHI K
Assistant Professor
Department of Electronics and Communication Engineering
JNN College of Engineering, Shimoga- 577204
Batch 21 2025-2026
 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Presentation Flow
• Introduction
• Literature survey
• Aim of the Project
• Objectives
• Block diagram
• Methodology
• Hardware/Software requirement
• Expected Results
• Action Plan
• References
 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Introduction
• Optimizing a neural network for identifying scene images involves improving its
accuracy, efficiency, and generalization capability.
• Convolution Neural Networks is one of the deep learning algorithms which
implemented to sample produces better outcomes as compared to other
machine learning algorithms.
• Deep Learning is one of the technique for predicting values.

• The data-set has been taken with some images from different types of location.
 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Introduction

• The project comprises several key components, including preprocessing

techniques to enhance image quality, feature extraction using CNNs to capture
relevant patterns, and classification algorithms to differentiate between different
types of images.

• By leveraging large-scale datasets and deep learning techniques, CNN-based

models can achieve state-of-the-art accuracy in identifying various environments
with minimal human intervention.
 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Introduction

1
 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Literature survey
Paper 1 : Robust Indoor and Outdoor classification of a Scene.

Authors, Methodology Results obtained Advantages Disadvantages

Year of publication,
type of journal

R.Raja, • Principal This paper has • Robust Feature • Computational Cost

S.Md. Mansoor Roomi, Component proposed a method for Extraction • Limited to Binary
D.Dharmalakshmi. Analysis (PCA) for automatic scene • High Accuracy: Classification
(Published on 2015) Feature Extraction categorization that Achieves 83.8% • Fixed Classifier
• Sparse finds application in the accuracy
Representation image retrieval • Noise Resilience
Classifier (SRC) to system.These features
categorize the are invariant to
image into indoor/ illumination, rotation
outdoor scene. and scale.
 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Literature survey
Paper 2 : Automated analysis of Scene using Image Feature Extraction.

Authors, Methodology Results obtained Advantages Disadvantages

Year of publication,
type of journal

S. Omar Gilani, • Gist to category of This method allows fast • Human-Like Scene • Limitations in Scene
Mohsin Jamil, contextual and efficient scene Understanding Similarity
Zahra Fazal, features. categorization with few • Scalability • Dependence on
Muhammad Samran • PCA is a method errors. It allowed images • Dimensionality Global Features
Navid, Rabeil Sakina. used for of same category to Reduction
(Published on 2016) dimensionality be correctly clustered • High Accuracy
reduction. and very few images,
• Clustering is used farthest from thecentroid
to classify similar of the clusters were
data items into wrongly categorized.
groups.
 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Literature survey
Paper 3 : Alex-Net model based Scene image classification technique

Authors, Methodology Results obtained Advantages Disadvantages

Year of publication,
type of journal

Jing Sun & Xibiao Cai, • Alex net In this paper, we • High Classification • Computational
Fuming Sun & Jianguo • D CNN extract the last layer Performance Complexity
Zhang features of the deep • Efficiency in Training • Dependence on Data
(Published on 2016) convolutional neural • End-to-End Learning Augmentation
network for scene • Limited
classification by using Interpretability
the Alex-Net model in
the deep learning
framework.
 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Literature survey
Paper 4 : Computer network connection enhancement optimization algorithm based on CNN
Authors, Methodology Results obtained Advantages Disadvantages
Year of publication,
type of journal

Hanju Li Sichuan CONVOLUTIONAL This proposed CNN- • Enhanced Feature • High Computational
University, Chengdu. NEURAL NETWORK based algorithm Extraction Capability Complexity
(Published on 2021) (CNN) enhances computer • Improved • Dependence on
network connections by Generalization Training Data Quality
fusing multiple CNNs, Performance • Limited Comparison
improving stability and • Dynamic with Other Methods
recognition accuracy. Optimization
The average field
neural network
algorithm proves
efficient and realistic for
optimization, offering
good economic
feasibility.
 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Literature survey
Paper 5 : Research on Image Classification Improvement Based on Convolutional Neural Networks with Mixed Training

Authors, Methodology Results obtained Advantages Disadvantages

Year of publication,
type of journal

Zhong-Qiu Zhao. • VGG16, Deep learning-based • Improved • Increased Complexity

(Published on 2022) ResNet152, and object detection excels Classification • Training Dependency
DenseNet, for in handling occlusion, Accuracy • Resource Intensive
image classification scale changes, and • Dynamic Weight
training in CNN background variations, Allocation
making it a research • Feature Diversity
hotspot. This paper Utilization
reviews frameworks
addressing challenges
like occlusion, clutter,
and low resolution, with
modifications on R-
CNN.
 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Aim of the Project

• To optimise neural networks for efficient scene image analysis and to
maintain good accuracy.
 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Objectives
• To collect the dataset of images.
• To preprocess the collected dataset.
• To test and train the augmented dataset.
• To evaluate and deploy trained model into applications.
 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Block diagram

Fig: Block Diagram For working of neural network

 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Methodology
1. Data Collection -
Dataset: Total 2080 scene image dataset
Categories: Select 6 classes {Mountain, Beach, Desert, Pond, Forest, Farmland}

2. Preprocessing -
Preprocessing consist of resizing standardise images to a fixed size(128x128).
Augmentation: Apply transformations (rotation, flipping, zooming) to reduce
overfitting.
 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Methodology
• Model Architecture: Two CNN architectures are employed
VGG-16:
- 13 convolutional layers + 3 fully connected layers.
- Uses small 3×3 kernels and max-pooling for hierarchical feature extraction.

• Training :
Transfer Learning: Initialise with pre-trained weights
Loss Function: Categorical cross-entropy for multi-class classification
Regularization: To prevent overfitting.
 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Methodology
• Evaluation:
Evaluate the performance of the trained model on a held-out test set. This will
give an estimate of how well the model will perform on new data.
• Deployment:
Integrate the trained model to production
 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Software requirement
• Tensor flow - An open-source deep learning framework used for building and
training neural networks
• NumPy/SciPy - numerical operations.
• Python 2.18.0- development environment
 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Hardware requirement
• GPU - Minimum NVIDIA GTX 1080 (8GB VRAM) for small-scale training.
• CPU & RAM:
CPU: Intel i7/Xeon or AMD Ryzen 7.
RAM: 8GB+
• Storage: SSD
 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Result obtained:
Objective 1:
• Dataset collection from Kaggle, consist of 2080 images which belonging to 6
classes.
• The classes are
-Pond
-Beach
-Farmland
-Desert
-Mountain
-Forest
 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Result obtained:
Objectives 2:
PRE-PROCESSING STEPS
1. Initialisation:
• We taken the image size as 128x128, batch size=32, total batches=65,
epoche =25
2. Normalization:
• In normalization we take the image batch, image size and its RGB value
 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Result obtained:
3. Augmentation
• In augmentation horizontal and vertical flipping, rotating.
 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Result obtained:
Objective 3:
1.Training data with augmentation
• Rescale=1.0 / 255: Normalizes pixel values from the range [0, 255] to [0, 1]
• validation_split=0.2: Splits 20% of the dataset for validation and keeps 80% for training
• rotation_range=20: Randomly rotates images up to 20 degrees
• width_shift_range=0.2: Randomly shifts the image horizontally (left or right) by up to
20% of the image width
• height_shift_range=0.2: Randomly shifts the image vertically (up or down) by up to 20%
of the image height
 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Result obtained:
• shear_range=0.2:Applies a slanting (shear) transformation.
• zoom_range=0.2:Randomly zooms in or out by up to 20%.
• horizontal_flip=True: Randomly flips images left-to-right.

2. Model complier:
• Adam optimizer- It is an efficient and adaptive method that adjusts learning rates during
training to help the model converge faster and more reliably.
• Loss function- which measures how far off the model’s predictions are from the actual
class labels.
 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Obtained Results:
3. Training and Validation:
• 1664 images are taken for training
• 416 images taken for validation
• The below table shows the accuracy of parameters
PARAMETER PERCENTAGE
TRAIN ACCURACY 87.46%
TEST ACCURACY 86%
LOSS 36.55%
 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Obtained Results: Training and validation output

 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Conclusion
• Optimization of the neural networks to identify scene images is crucial to the development of
visual recognition automation systems. Application of a robust CNN design and appropriate
preprocessing of data, augmentation techniques, and hyperparameters such as the learning
rate, batch size, and number of epochs can help improve the performance of the model
significantly. Methods like data normalization, image data augmentation, and dropout are
utilized to prevent overfitting and improve generalizability across multiple environments.
• The application of the Adam optimizer with the categorical cross-entropy loss function,
coupled with a well-selected activation strategy (e.g., ReLU and Softmax), enables effective
training and correct multi-class classification. The accuracy metrics illustrate that with correct
optimization, the model can effectively classify multiple scene classes (e.g., beach, mountain,
forest) even under complicated visual scenarios.
• This is done through emphasizing the capability of deep learning to comprehend scenes at
the same time as providing a firm groundwork for its application to autonomous systems,
intelligent surveillance, and image retrieval based on content.
 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

Action Plan:
Sl. No. Tasks to be performed Deadline ( Expected date/week of completion)

1. Literature Survey 24/03/2025

2. Synopsis Submission 01/04/2025

3. Submission of First Phase Report 06/05/2025

4. Objective 1 30/03/2025

5. Objective 2 10/04/2025

6. Objective 3 04/05/2025

7. Objective 4

8. Testing & Result Analysis

9. Documentation of Project Report

 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

References
1. Szegedy, Christian, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent
Vanhoucke, and Andrew Rabinovich. (2015) "Going deeper with convolutions." In Proceedings of the IEEE conference on
computer vision and pattern recognition
2. Gao, Jingyu, Jinfu Yang, Guanghui Wang, and Mingai Li. (2016) "A novel feature extraction method for scene recognition
based on centred convolutional restricted Boltzmann machines." Neuro computing 214: 708-717.
3. Masood, Sarfaraz, Tarun Luthra, Himanshu Sundriyal, and Mumtaz Ahmed. (2017) "Identification of diabetic retinopathy
in eye images using transfer learning." In 2017 International Conference on Computing, Communication and Automation
(ICCCA), pp. 1183-1187. IEEE.
4. Kibria, Sakib B., and Mohammad S. Hasan. (2017) "An analysis of Feature extraction and Classification Algorithms for
Dangerous Object Detection." In 2017 2nd International Conference on Electrical & Electronic Engineering (ICEEE), pp. 1-
4. IEEE.
5. D. Jia, D. Wei, L.J. Li, L. Kai, and F.F. Li, ImageNet: A Large-scale Hierarchical Image Database, IEEE Conference on
Computer Vision and Pattern Recognition, 2009.
6. Y. Jiang, Research on scene image content representation and classification, National University of Defence Technology
Doctoral Thesis, 2010.
 National Education Society®
J N N College of Engineering, Shivamogga
Department of Electronics & Communication Engineering

THANK YOU
*