Face Recognition System using OpenCV (Classical CV)

A real-time face recognition system built with classical computer vision — Haar Cascade for face detection and LBPH (Local Binary Pattern Histograms) for face recognition. No deep learning; this project demonstrates the classical CV pipeline end-to-end and its real-world trade-offs.

What it does

• Collects face images from a webcam and labels them
• Trains an LBPH recognizer on the collected faces
• Runs live recognition, labeling known faces and flagging unknown ones
• Prints a confidence (distance) score per frame for monitoring

Detection vs. Recognition

• Detection (Haar Cascade): finds where a face is in the frame.
• Recognition (LBPH): identifies whose face it is.

Tech stack

• Python 3
• opencv-contrib-python (the contrib build is required for cv2.face)
• NumPy

Setup

python -m venv .venv
.venv\Scripts\Activate.ps1        # Windows PowerShell
pip install -r requirements.txt

Note: install opencv-contrib-python, not opencv-python — the LBPH recognizer (cv2.face.LBPHFaceRecognizer_create) lives in the contrib module.

How to run (in order)

1. python 01_collect_faces.py — capture ~100 face crops into datasets/
2. python 02_train_model.py — train LBPH, produces models/face_trained.yml
3. python 03_recognize_live.py — live recognition with name labels

Press q to quit any webcam window.

Understanding the confidence score

LBPH confidence is a distance, not a percentage — lower means a better match. Typical values for a correctly recognized face are ~20–40; the script labels a face as "Unknown" when the distance exceeds the threshold.

Known limitations (and why)

• Lighting sensitivity: LBPH compares raw pixel-intensity texture, so a lighting change between training and testing raises the distance and degrades accuracy.
• Scale sensitivity: trained on faces at one rough size, so recognition is best at a similar distance from the camera.
• Per-frame flicker: each frame is classified independently with no temporal smoothing, so borderline matches flip between the name and "Unknown."

Possible improvements

• Smooth predictions across frames (majority vote over N frames)
• Replace Haar with a DNN face detector (more robust to angle/lighting)
• Replace LBPH with face embeddings (FaceNet / dlib) for far better accuracy
• Multi-person support with a label→name mapping file and attendance logging

A second experiment: celebrity_demo/

A parallel version that recognizes a small set of public figures from static image folders, using the same Haar + LBPH approach. Included to contrast recognition on curated webcam data vs. uncontrolled internet images — the latter produces much weaker matches, illustrating LBPH's data-quality dependence.