How Deep Learning Transforms Modern Face Recognition: From Basics to DeepFace

Traditional face recognition involved image capture, preprocessing, identity verification, and search. Modern systems extend to driver detection in ADAS, pedestrian tracking, and dynamic object tracking, moving from simple image processing to real‑time video processing, with algorithms shifting from AdaBoost and PCA to CNN, R‑CNN, and 3‑D methods.

Current research focuses on deep‑learning approaches for video face recognition.

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

Main open‑source projects :

SeetaFace engine – C++ implementation, BSD‑2 license, developed by Prof. Shan’s team.

Major APIs/SDKs :

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

Skybiometry – provides face detection, recognition, and grouping.

Popular face image datasets : LFW (Labelled Faces in the Wild) and YFW (YouTube Faces in the Wild). LFW is widely used; current image‑based recognition accuracy exceeds 99%.

Leading Chinese companies in face recognition include Hanwang Technology, iFlytek (with Gaussian Face achieving 98.52% on LFW and DeepID2 reaching 99.4%), ZhiSheng (3‑D face recognition), and SenseTime (106 facial keypoints, deep‑learning focus).

Face Recognition Process

Four stages: face detection, face alignment, face verification, and face identification.

Face detection : Detect faces and draw bounding boxes; OpenCV’s Haar cascade is a classic method.

Face alignment : Correct pose using 2D or 3D alignment based on fiducial points (eyes, nose, mouth corners). Images are normalized to a fixed size.

Face verification : Pair‑matching to decide same or different person, often used in small‑office access control.

Face identification : Classify the aligned face to a known identity.

Face Recognition Categories

Three main categories: image‑based, video‑based, and 3‑D face recognition.

Image‑based : Traditional methods (PCA, SVM, Bayesian, etc.) and modern deep learning (DeepFace, Face++).

Video‑based : Real‑time tracking and temporal association.

DeepFace

DeepFace, proposed by Facebook, pioneered CNN‑based face recognition. It uses a 152×152 input, several convolutional and locally connected layers, followed by two fully‑connected layers (F7, F8) and a softmax classifier. Features are L2‑normalized; training minimizes cross‑entropy loss via SGD.

Key architecture details:

Conv1: 32 filters of 11×11×3.

Max‑pooling: 3×3 stride 2.

Conv2: 16 filters of 9×9.

Local‑Conv layers with non‑shared weights (9×9, 7×7, 5×5).

Fully‑connected: 4096‑dim.

Softmax: 4030‑dim.

Training uses large datasets (e.g., Social Face Classification 4.4 M faces, LFW, YTF). Results on LFW and YTF are shown.

Verification Methods

Chi‑square distance is used to compare normalized DeepFace vectors, which are sparse and lie in [0, 1].

Siamese networks perform end‑to‑end metric learning by feeding two face embeddings into a fully‑connected layer that outputs a same/different decision.

Installation and Usage

Supported on Python 3/2, tested on macOS and Linux. Install via pip: pip3 install face_recognition Example Python code:

import face_recognition

image = face_recognition.load_image_file("your_file.jpg")

face_locations = face_recognition.face_locations(image)

The command‑line tool face_recognition can identify faces in a single image or a folder, outputting filenames and recognized names.

For detailed API documentation see https://face-recognition.readthedocs.io .