Face Detection with SVM and HOG Features

Binary face detection system built with classical computer vision techniques - no deep learning required.

Overview

The model classifies image regions as face or non-face using HOG (Histogram of Oriented Gradients) features fed into a Support Vector Machine (SVM) with linear kernel. A sliding window approach enables detection on full images at multiple scales.

Dataset

Split	Source	Size
Faces	Labeled Faces in the Wild (LFW) - Kaggle	5,000 images
Non-faces	COCO 2017 Train (filtered: no person category)	5,000 images

Non-face images were filtered programmatically using COCO JSON annotations to exclude any image containing a person, ensuring clean negative examples.

Pipeline

1. Preprocessing - grayscale conversion, resize to 150×150, normalization to [0, 1]
2. Augmentation - random rotation, scaling, translation and brightness variation (batch processing to control RAM usage)
3. Feature extraction - HOG features via scikit-image
4. Training - SVM with linear kernel, optimized with GridSearchCV
5. Inference - sliding window at multiple scales (75, 50, 25 px) with probability thresholding (default: 0.65)

Results

Evaluated on 30% hold-out test set (3,000 samples).

Metric	Score
ROC-AUC	0.9597
Average Precision (AP)	0.9575
Correct predictions (TP + TN)	2,683 / 3,000 (89.4%)
False negatives (missed faces)	174
False positives (non-faces misclassified)	143

Tech Stack

Python · NumPy · OpenCV · scikit-image · scikit-learn · Matplotlib · joblib

How to Run

The notebook is optimized for Google Colab.

1. Open face_detection.ipynb on Google Colab
2. Run cells sequentially
3. The trained model is saved automatically to Google Drive as face_detection_svm.pkl

To run inference on a new image, load the saved model:

import joblib
model = joblib.load("face_detection_svm.pkl")

Project Structure

face_detection/
└── face_detection.ipynb   # Main notebook (training + inference)