[Paper Review] Face Recognition: From Traditional to Deep Learning Methods

Daniel Sáez-Trigueros, Li Meng+1 more|arXiv (Cornell University)|Oct 31, 2018

Face recognition and analysis101 references104 citations

TL;DR

The paper surveys face recognition from traditional geometry, holistic, feature-based, and hybrid methods to modern deep learning approaches, detailing building blocks, datasets, and representative methods and performance.

ABSTRACT

Starting in the seventies, face recognition has become one of the most researched topics in computer vision and biometrics. Traditional methods based on hand-crafted features and traditional machine learning techniques have recently been superseded by deep neural networks trained with very large datasets. In this paper we provide a comprehensive and up-to-date literature review of popular face recognition methods including both traditional (geometry-based, holistic, feature-based and hybrid methods) and deep learning methods.

Motivation & Objective

• Survey the evolution of face recognition methods from geometry-based to deep learning.
• Compare holistic, feature-based, and hybrid approaches and their robustness to real-world variations.
• Explain common building blocks of face recognition systems (detection, alignment, representation, matching).
• Highlight large-scale datasets enabling CNN-based performance improvements.
• Summarize representative methods and their reported accuracies on standard benchmarks.

Proposed method

• Classify methods into geometry-based, holistic, feature-based, hybrid, and deep learning categories.
• Describe PCA/LDA, kernel methods, sparse representations, and SVMs as foundational techniques.
• Explain face detection, alignment, representation, and matching building blocks.
• Discuss CNN-based strategies including classification-based training, bottleneck features, and metric learning (e.g., Siamese, contrastive loss).
• Cite landmark systems (e.g., DeepFace, DeepID) and their architectural innovations and alignment strategies.
• Summarize performance on benchmarks such as LFW with reported accuracies.

Experimental results

Research questions

• RQ1 What are the main historical categories of face recognition methods and their strengths/limitations?
• RQ2 How have deep learning approaches improved face recognition compared to traditional methods?
• RQ3 What building blocks are essential for effective face recognition systems, and how do they interact?
• RQ4 What datasets and architectures have driven performance gains on standard benchmarks like LFW?

Key findings

• Holistic methods using PCA/LDA and sparse representations achieved notable accuracy prior todeep learning, with reported LFW performance such as 92.4% for a holistic approach.
• CNN-based face recognition dramatically improved accuracy when trained on large-scale in-the-wild datasets, exemplified by DeepFace achieving 97.35% on LFW.
• The GaussianFace approach achieved 98.52% on LFW, illustrating strong cross-domain discriminative capability.
• Hybrid methods combining local features (LBP, Gabor, SIFT) with subspace learning often outperformed single-technique approaches.
• Earlier CNN-based methods required large datasets and higher-capacity networks to surpass traditional techniques, setting the stage for current state-of-the-art results.