UAV-Attention-GradCAM-Lite

A lightweight PyTorch implementation demonstrating Spatial Attention mechanisms and Grad-CAM interpretability for UAV imagery analysis. Built as an open-source demonstration of architectural competencies used in commercial computer vision projects.

Visual Results

Grad-CAM Heatmap Visualization

Original Image	Grad-CAM Heatmap

Additional Examples

Original Image	Grad-CAM Heatmap

The heatmaps highlight regions of the image that the model focuses on during inference, providing interpretability for the attention mechanism.

Overview

This project implements:

• CBAM-style Spatial Attention: Integrated after layer4 in a pre-trained ResNet18 backbone
• Grad-CAM Visualization: Generates class activation maps from attention feature maps
• UAV Image Processing: Optimized for aerial imagery analysis

Repository Structure

UAV-Attention-GradCAM-Lite/
├── assets/                    # Input images and generated heatmaps
├── spatial_attention.py       # Spatial attention module implementation
├── gradcam_engine.py         # Grad-CAM computation and visualization
├── requirements.txt          # Python dependencies
└── README.md                 # Documentation

Installation

Prerequisites

• Python 3.7+
• pip or conda package manager

Setup

1. Clone the repository:

git clone https://github.com/HoneyBadger-010/UAV-GradCAM-Lite.git
cd UAV-GradCAM-Lite

1. Create a virtual environment (recommended):

python3 -m venv .venv
source .venv/bin/activate  # On Windows: .venv\Scripts\activate

1. Install dependencies:

pip install --extra-index-url https://download.pytorch.org/whl/cpu torch torchvision opencv-python numpy matplotlib

Or use the requirements file:

pip install -r requirements.txt

Usage

Run the Grad-CAM engine on images in the assets/ directory:

python gradcam_engine.py

The script will:

1. Load a pre-trained ResNet18 model with spatial attention
2. Process all .jpg and .png images in ./assets/
3. Generate Grad-CAM heatmap overlays
4. Save results with _heatmap suffix

Example Output

For an input file assets/example.jpg, the script generates:

assets/example_heatmap.jpg

How It Works

Spatial Attention Module

The spatial_attention.py module implements a CBAM-inspired spatial attention mechanism:

• Applies max and average pooling across the channel dimension
• Generates spatial attention maps through a convolutional layer
• Multiplies attention weights with input features

Grad-CAM Engine

The gradcam_engine.py script:

1. Preprocesses images (resize, normalize with ImageNet statistics)
2. Performs forward pass through the attention-enhanced ResNet
3. Computes gradients with respect to the predicted class
4. Generates class activation maps from attention features
5. Overlays heatmaps on original images using OpenCV

Technical Details

• Model: ResNet18 with spatial attention after layer4
• Input Size: 224x224 pixels
• Normalization: ImageNet mean and standard deviation
• Heatmap: ReLU-activated weighted combination of feature maps
• Overlay: 50% alpha blending with JET colormap

Applications

• UAV/drone imagery analysis
• Object detection interpretability
• Attention mechanism visualization
• Model debugging and validation

License

This project is open-source and available for educational and research purposes.

Acknowledgments

• PyTorch team for the deep learning framework
• ResNet architecture from "Deep Residual Learning for Image Recognition"
• CBAM attention mechanism from "Convolutional Block Attention Module"
• Grad-CAM from "Grad-CAM: Visual Explanations from Deep Networks"