Comprehensive Overview of Face Recognition Techniques and DeepFace Implementation

Traditional face recognition involved image acquisition, preprocessing, identity confirmation, and search, but recent advances have extended it to driver monitoring in ADAS, pedestrian tracking, and dynamic object tracking.

The field has shifted from statistical methods such as AdaBoost and PCA to deep learning approaches like CNN and RCNN, with growing interest in 3‑D face recognition supported by academia, industry, and government.

Key researchers : Prof. Shan Shiguang (Institute of Computing Technology, CAS), Prof. Li Ziqing (CAS Institute of Biometrics), Prof. Su Guangda (Tsinghua University), Prof. Tang Xiaoe (CUHK), Ross B. Girshick, among others.

Main open‑source projects :

SeetaFace engine – a BSD‑2 licensed C++ face‑recognition library without external dependencies.

Major APIs/SDKs :

Face++ – cloud service offering detection, recognition, and attribute analysis.

SkyBiometry – provides detection, recognition, and grouping.

Popular face image datasets : LFW (Labeled Faces in the Wild) and YFW (YouTube Faces in the Wild); LFW is widely used and current algorithms achieve ~99% accuracy.

Chinese companies and their focus :

Hanwang Technology – identity verification for access control.

iFlytek – Gaussian Face method (98.52% on LFW, later DEEPID2 reached 99.4%).

ChuanDa Zhisheng – 3‑D face recognition and camera commercialization.

SenseTime – deep‑learning‑driven solutions across face, text, body, vehicle, and object recognition, with 106 facial key‑points detection.

Face recognition pipeline consists of four modules:

Face detection – locating faces, e.g., using OpenCV Haar cascades.

Face alignment – correcting pose via 2‑D or 3‑D landmark‑based transformations.

Face verification – pair‑matching to decide if two faces belong to the same person.

Face identification – classifying a face into one of many known identities.

DeepFace (Facebook) introduced a CNN‑based framework that later inspired DeepID and FaceNet. Its basic workflow is:

face detection → face alignment → face verification → face identification

DeepFace architecture (input 152×152 after 3‑D alignment):

Conv1: 32 filters of size 11×11×3

Max‑pooling: 3×3 stride 2

Conv2: 16 filters of size 9×9

Local‑Conv layers (non‑shared weights) of sizes 9×9, 7×7, 5×5

Fully‑connected layer: 4096 units

Softmax: 4030 units

Low‑level features are extracted through the first three layers (C1, M2, C3). Subsequent local‑connection layers capture region‑specific patterns, and the final fully‑connected layers (F7, F8) produce a 4096‑dimensional face descriptor, which is L2‑normalized.

Training minimizes cross‑entropy loss using stochastic gradient descent; ReLU activation yields sparse features (>75% zeros) and dropout is applied only to F7.

Evaluation on datasets:

Social Face Classification (4.4M faces, 4030 identities)

LFW (13,323 faces, 5,749 identities)

YouTube Faces (3,425 videos, 1,595 identities)

Results: DeepFace achieves 99.38% accuracy on LFW, comparable to state‑of‑the‑art methods.

Practical usage with the face_recognition Python library (requires Python 3/2, tested on macOS/Linux):

import face_recognition</code>
<code>image = face_recognition.load_image_file("your_file.jpg")</code>
<code>face_locations = face_recognition.face_locations(image)

After installing dlib from source and pip install face_recognition, the command‑line tool face_recognition can identify faces in a folder given a reference folder of labeled images.

Overall, the article consolidates the theoretical background, algorithmic details, datasets, and hands‑on tools for modern face recognition research and development.