Face Detection with a Lightweight CNN

Overview

This project implements a lightweight convolutional neural network (CNN) for binary face classification (face vs noface) and a sliding-window detection pipeline. The detector builds an image pyramid, scans with a fixed-size window, classifies each window, and applies Non-Maximum Suppression (NMS) to produce final bounding boxes. The repo includes training scripts, inference scripts, a detection entrypoint, a submission-ready report notebook, a simply web app for face detection, and demo images with results.

Repository Structure

• train.py — training script.
• test.py — test script.
• predict.py — classification inference for single image/patch.
• face_detection.py — sliding-window detection entrypoint.
• net.py — CNN model definition (PyTorch).
• load_data.py — dataset loader and preprocessing.
• app.py — Streamlit web app for face detection.
• model.pth — model checkpoint.
• face_detection_report.ipynb — Jupyter report for submission.
• demo_images/ — demo inputs and detection outputs (*_result.jpg).
• results/ — training/validation accuracy charts (train_accuracy.png, validation_accuracy.png).
• utils/gif2jpg.py — utility for GIF → JPG.
• torchsampler/ — imbalanced sampler helpers.

Installation

• Install dependencies:

pip install -r requirements.txt

• If you encounter NumPy/PyTorch compatibility errors (e.g., “compiled against NumPy 1.x” or “Numpy is not available”), align versions by choosing ONE of the following:

pip install 'numpy<2'

pip install --upgrade torch torchvision torchaudio

Quick Start

• Train (optional):

python train.py

• Test the model on the test set:

python test.py

• Classify a single image/patch:

python predict.py /path/to/image.jpg

• Detect faces on a full image:

python face_detection.py /path/to/image.jpg

• Run the Streamlit web app:

streamlit run app.py

Model and Preprocessing

• Task: Binary classification (face vs noface), used as the backbone for detection.
• Architecture: Compact CNN with convolution, ReLU, pooling, and fully connected layers ending in 2 logits for CrossEntropyLoss.
• Input Size Consistency:
  • Keep inference resize consistent with training (e.g., 36×36) to avoid fc1 input dimension mismatches.

Training

• Command:

python train.py

• Typical setup:
  • Loss: CrossEntropyLoss
  • Optimizer: Adam (e.g., learning rate 1e-3)
  • Logging: print per-epoch loss/accuracy; log to TensorBoard
• Checkpoints:
  • Saved to model.pth. Ensure inference uses the same preprocessing as training.

Detection Pipeline

• Steps:
  • Image Pyramid → Sliding Window → CNN Classification → NMS
• Thresholds:
  • Tune the probability threshold to balance Precision vs Recall.
• Output:
  • Bounding boxes overlaid on the input image.

License

This project is intended for learning and experimentation. No specific license is provided.