Bird Species Classification Using CNNs

Overview

This project explores a convolutional neural network (CNN)-based approach for Bird Species Classification. The model's performance is analyzed across various optimization techniques, providing insights into its learning behavior and interpretability through Class Activation Maps (CAM).

---

Features

1. Model Architecture:

   • A robust CNN architecture featuring residual connections, Batch Normalization, and Adaptive Average Pooling for efficient representation learning.

2. Training Insights:

   • Performance metrics, including training and validation loss/accuracy trends across epochs, are visualized for clear understanding.

3. Optimization Techniques:

   • Impact of methods such as data augmentation, dropout, Compound Scaling Laws, and StepLR scheduling on model accuracy is explored and tabulated.

4. Class Activation Maps (CAM):

   • Visualization of regions focused on by the model for classification decisions, providing interpretability and identifying areas of misclassification.

---

Model Architecture

Layer	Configuration	Output Shape
Input	Image (300x300x3)	(300, 300, 3)
Conv2D + BatchNorm + ReLU	7x7, stride 2, padding 3	(64, 150, 150)
MaxPool2D	3x3, stride 2	(64, 75, 75)
Residual Blocks	Several Conv2D layers with skip connections and strides	Intermediate (128 to 512 channels)
AdaptiveAvgPool2D	Adaptive Average Pooling	(512, 1, 1)
Fully Connected Layer	Linear (512 → 10)	(10)

The table captures the architecture of the birdClassifier CNN, optimized for scalability and feature extraction.

---

Performance Metrics

Training and Validation Trends

Plots for Loss and Accuracy across epochs highlight model convergence and generalization:

• Training-to-validation split: 80% train, 20% validation.

Optimization Techniques

Technique	Validation Accuracy	Notes
No Optimization	60.13%	Baseline model
Compound Scaling Laws	81.78%	Improved scaling
Data Augmentation	89.98%	Enhanced generalization with augmented data
Dropout	86.50%	Slower learning but robust performance
StepLR (all combined)	90.72%	Best accuracy achieved

---

Class Activation Maps (CAM)

CAM visualizations provide insights into the model's decision-making:

• Heatmaps reveal the image regions most influential in classification.
• Misclassified examples show scattered attention, indicating potential areas for further optimization.

Examples of CAM:

• Correct classifications emphasize relevant textures and objects in the images.

• Misclassified examples display poor focus, leading to incorrect predictions.

---

Usage

1. Clone the repository:

   git clone <repository-url>
   cd <repository-directory>

2. Running the Code

   # Running code for training. save the model in the same directory with name "bird.pth"
   python bird.py path_to_dataset train bird.pth 
   
   # Running code for inference
   python bird.py path_to_dataset test bird.pth

---

Authors

• Karan Deo Burnwal

---

Acknowledgments

This work was completed as part of COL333 Assignment 3.1 at IIT Delhi, exploring advanced deep learning techniques under academic supervision.