eXCNN: Explainable CNN Visualization Platform

eXCNN is a comprehensive interactive platform designed to visualize and decode the decision-making process of Convolutional Neural Networks (CNNs). By leveraging state-of-the-art Explainable AI (XAI) methods, it allows users to see exactly "where" and "what" a deep learning model looks at when classifying images.

Key Features

• Interactive Analysis: Upload any image (JPG/PNG) to receive real-time classification and visual explanations.
• Multi-Method Visualization:
  • Grad-CAM: Visualizes class-discriminative regions using gradient information.
  • Occlusion Sensitivity: Measures feature importance by systematically hiding image parts.
  • Guided Backpropagation: Highlights fine-grained pixel-level details.
  • Guided Grad-CAM: Combines localization with high-resolution details for the sharpest insight.
• Precomputed Gallery: Explore curated examples with detailed breakdowns of model behavior.
• Robust Backend: Powered by PyTorch models (ResNet50) and FastAPI for high-performance inference.
• Modern Frontend: A clean, monochrome interface focusing on usability and data visualization.

Interface Gallery

Analysis View
Feature Visualization 1	Feature Visualization 2

---

Technology Stack

• Backend: Python 3.10+, FastAPI, PyTorch, Torchvision, NumPy, OpenCV, Uvicorn.
• Frontend: Native HTML5, CSS3 (Modern/Flexbox/Grid), Vanilla JavaScript (ES6+).
• Model: ResNet50 (Pre-trained on ImageNet-1k).

Installation & Setup

Prerequisites

• Python 3.8 or higher
• pip (Python package manager)

1. Clone & Setup

Clone the repository and install the dependencies:

# Install required Python packages
pip install -r requirements.txt

2. Run the Application

Start the entire platform with a single command:

python run_app.py

This will automatically launch:

• Backend API at http://localhost:8000
• Frontend Interface at http://localhost:8080

Wait for the "[OK] Application is fully running!" message, then open your browser.

3. Access

Open your browser and navigate to: http://localhost:8080

---

Project Structure

eXCNN/
├── assets/                 # Static assets and database files
│   ├── imagenet_labels.json # Full 1000 ImageNet class labels
│   ├── screenshots/        # Application screenshots
│   └── precomputed/        # Gallery examples with generated visualizations
├── backend/                # FastAPI application
│   ├── api.py              # Main API endpoints
│   ├── explainers.py       # Implementation of XAI methods (GradCAM, etc.)
│   └── model_loader.py     # ResNet50 model management
├── frontend/               # User Interface
│   ├── index.html          # Main application structure
│   ├── styles.css          # Monochrome design system
│   └── app.js              # Application logic and API integration
├── scripts/                # Utility scripts
│   ├── generate_precomputed.py # Script to generate gallery assets
│   └── test_backend.py     # Backend test suite
├── requirements.txt        # Python dependencies
└── README.md               # This documentation

---

Understand the XAI Methods

Method	Description	Best For
Grad-CAM	Uses the gradients of the target class flowing into the final convolutional layer to produce a coarse localization map highlighting important regions.	identifying where the object is.
Occlusion	Slides a grey patch over the image and records the drop in probability. If confidence drops drastically, that area is critical.	Verifying robustness without gradients.
Guided Backprop	visualizes gradients with respect to the image, but zeroing out negative gradients during backpropagation to show only positive influences.	Seeing textures and patterns.
Guided Grad-CAM	Element-wise multiplication of Guided Backprop and Grad-CAM (bilinear upsampled).	The most interpretable and sharp visualization.

---

License

This project is built for educational purposes. Code is open for modification and study.