CIFAR-100 Image Classification using CNN

This project involves building and optimizing a Convolutional Neural Network (CNN) to classify images from the CIFAR-100 dataset into 100 fine-grained categories. The model leverages advanced deep learning techniques, such as SE (Squeeze-and-Excitation) blocks, Batch Normalization, and Data Augmentation, to enhance classification accuracy.

---

Table of Contents

• Problem Statement
• Video Demo
• Dataset Overview
• Model Architecture
• Features and Techniques
• Results
• Setup and Usage
• Saved Model Integration
• Future Improvements
• Contributions
• License

---

Problem Statement

The CIFAR-100 dataset comprises 32x32 color images across 100 categories, ranging from animals and plants to everyday objects. The goal of this project is to classify these images using a robust and optimized deep learning model.

Key objectives include:

• Designing a scalable CNN architecture.
• Using SE blocks for channel recalibration.
• Implementing data augmentation for better generalization.
• Achieving improved accuracy using batch normalization and dropout.

---

Video Demo

Screen.Recording.2024-12-18.at.2.14.18.PM.mov

---

Dataset Overview

• Training Samples: 50,000 images
• Test Samples: 10,000 images
• Image Dimensions: 32x32 pixels with 3 color channels (RGB)
• Classes: 100 distinct categories, such as 'apple', 'bicycle', 'lion', etc.

The dataset is publicly available and a standard benchmark for image classification tasks.

---

Model Architecture

The CNN architecture designed for this project includes:

1. Convolutional Layers: For feature extraction, with increasing complexity in subsequent layers.
2. SE (Squeeze-and-Excitation) Blocks: To dynamically recalibrate channel importance.
3. Batch Normalization: For faster convergence and stable training.
4. Pooling Layers: MaxPooling layers to reduce spatial dimensions.
5. Fully Connected Layers: Dense layers for classification.
6. Dropout Layers: For regularization to prevent overfitting.

---

Features and Techniques

1. Data Augmentation: Random transformations (flipping, rotation) applied to training data to enhance diversity.
2. SE Blocks: A feature recalibration mechanism to improve learning of important features.
3. Batch Normalization: Standardizes activations in intermediate layers for better training dynamics.
4. Learning Rate Scheduling: An exponential decay schedule for optimal training convergence.
5. Saved Model Format: The trained model is saved in .keras format for deployment readiness.

---

Results

The final model achieved the following metrics on the CIFAR-100 dataset:

• Test Accuracy: ~66%
• Test Loss: ~1.64

Visualization

Sample predictions are visualized, showcasing the model's confidence in the assigned class labels.

---

Setup and Usage

Prerequisites

Ensure you have the following installed:

• Python 3.8+
• TensorFlow 2.0+

Steps to Run the Project

1. Clone the repository:

   git clone https://github.com/obiwan04kanobi/CNN_CIFAR-100.git
   cd CNN_CIFAR-100

2. Test the saved model with an image:

   from tensorflow.keras.models import load_model
   model = load_model('saved_models/cifar100_67_model.keras')

Predicting on a New Image

Save an image in the project directory and run:

from tensorflow.keras.preprocessing import image
img = image.load_img('your_image.png', target_size=(32, 32))
img_array = image.img_to_array(img) / 255.0
img_array = img_array[np.newaxis, ...]
prediction = model.predict(img_array)
print(f"Predicted class: {classes[np.argmax(prediction)]}")

---

Saved Model Integration

The trained .keras model can be easily integrated into:

1. Django/Flask Web Applications: Use TensorFlow's APIs to load the model and serve predictions.
2. Mobile Applications: Convert the model to TensorFlow Lite (.tflite) for efficient on-device inference.
3. Cloud Deployment: Deploy the model as a REST API using platforms like AWS, Azure, or GCP.

---

Future Improvements

1. Accuracy Optimization: Experiment with deeper networks or pre-trained models (transfer learning).
2. Hyperparameter Tuning: Fine-tune learning rates, dropout rates, and layer configurations.
3. Custom Loss Functions: Tailor loss functions to improve classification performance.
4. Interactive Web Application: Build a live web app for real-time predictions.

---

License

This project is licensed under the MIT License. See the LICENSE file for details.

---

Acknowledgments

• CIFAR-100 Dataset: Krizhevsky, 2009
• Squeeze-and-Excitation Networks: Hu et al., 2018
• TensorFlow and Keras community for providing tools to build and evaluate the model.