🔍 VisionInsight: Explainable Pet Classifier

🌟 Overview

VisionInsight is an advanced image classification system that not only identifies cats and dogs but also explains its decision-making process through visual saliency maps. This project demonstrates the practical application of Explainable AI (XAI) techniques in deep learning, making neural network decisions more transparent and interpretable.

🎯 Features

• Real-time cat and dog classification using ResNet50
• Interactive web interface powered by Gradio
• Visual explanation through saliency maps
• Confidence score visualization
• Easy deployment in Google Colab

🧠 Explainable AI Components

This project implements several key concepts in Explainable AI:

Saliency Maps

Saliency maps highlight the regions of an input image that most strongly influence the model's classification decision. Our implementation uses gradient-based visualization, which:

• Computes gradients of the output with respect to the input image
• Identifies pixels that would most significantly affect the classification if changed
• Visualizes these important regions using a blue-scale heatmap

Confidence Scores

The system provides detailed confidence scores for both categories (cats and dogs), offering:

• Normalized probability distributions
• Clear percentage-based confidence metrics
• Visual representation of decision certainty

🛠️ Technical Implementation

Model Architecture

• Base Model: ResNet50 (pretrained on ImageNet)
• Input Processing: 224x224 image size with normalization
• Output: Binary classification (Cat vs Dog) with confidence scores

Key Components

1. Image Preprocessing

   • Resize to standard dimensions
   • Normalize using ImageNet statistics
   • Convert to PyTorch tensors

2. Classification Pipeline

   • Forward pass through ResNet50
   • Probability computation using softmax
   • Class-specific confidence calculation

3. Saliency Computation

   • Gradient computation w.r.t input
   • Gradient pooling across channels
   • Normalization and visualization

🚀 Getting Started

Prerequisites

pip install torch torchvision gradio pillow numpy

Running the Application

python app.py

Google Colab Usage

1. Open the provided notebook
2. Run the installation cell
3. Execute the application cell
4. Click the generated public URL

📊 Example Usage

from pet_classifier import PetClassifier

# Initialize the classifier
classifier = PetClassifier()

# Classify an image
results = classifier.classify_image(image)

# Access results
original_image = results['original_image']
saliency_map = results['saliency_map']
prediction = results['prediction']
confidence = results['confidence']

🔧 Customization

You can customize various aspects of the system:

• Modify the confidence threshold
• Adjust saliency map coloring
• Change the model architecture
• Add support for additional classes

🧪 Technical Details

Saliency Map Generation

The saliency maps are generated using the following process:

def compute_saliency_map(input_tensor, target_class):
    input_tensor.requires_grad_()
    output = model(input_tensor)
    output[0, target_class].backward()
    gradients = input_tensor.grad.data.abs()
    saliency_map = torch.max(gradients, dim=1)[0]
    return saliency_map

Model Configuration

The system uses standard ImageNet normalization:

transforms.Normalize(
    mean=[0.485, 0.456, 0.406],
    std=[0.229, 0.224, 0.225]
)

📚 Further Reading

• Understanding Saliency Maps
• ResNet Architecture
• Explainable AI Techniques

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

📜 License

This project is licensed under the MIT License - see the LICENSE.md file for details.

✨ Acknowledgments

• The PyTorch team for their excellent deep learning framework
• The Gradio team for their user interface components
• The scientific community for advancing Explainable AI techniques

📧 Contact

For questions and feedback, please open an issue in the GitHub repository.