[Paper Review] Large-scale Supervised Hierarchical Feature Learning for Face Recognition

Jianguo Li, Yurong Chen|arXiv (Cornell University)|Jul 6, 2014

Face and Expression Recognition27 references3 citations

TL;DR

This paper proposes a large-scale supervised hierarchical feature learning framework for face recognition that operates in three stages: channel-level selection using Gaussian receptive fields, patch-level selection from an over-complete pool, and discriminant subspace projection of patch descriptors. The method achieves state-of-the-art performance on FRGC and LFW benchmarks with low memory and computational cost, enabling deployment on embedded systems.

ABSTRACT

This paper proposes a novel face recognition algorithm based on large-scale supervised hierarchical feature learning. The approach consists of two parts: hierarchical feature learning and large-scale model learning. The hierarchical feature learning searches feature in three levels of granularity in a supervised way. First, face images are modeled by receptive field theory, and the representation is an image with many channels of Gaussian receptive maps. We activate a few most distinguish channels by supervised learning. Second, the face image is further represented by patches of picked channels, and we search from the over-complete patch pool to activate only those most discriminant patches. Third, the feature descriptor of each patch is further projected to lower dimension subspace with discriminant subspace analysis. Learned feature of activated patches are concatenated to get a full face representation.A linear classifier is learned to separate face pairs from same subjects and different subjects. As the number of face pairs are extremely large, we introduce ADMM (alternative direction method of multipliers) to train the linear classifier on a computing cluster. Experiments show that more training samples will bring notable accuracy improvement. We conduct experiments on FRGC and LFW. Results show that the proposed approach outperforms existing algorithms under the same protocol notably. Besides, the proposed approach is small in memory footprint, and low in computing cost, which makes it suitable for embedded applications.

Motivation & Objective

• Address the challenge of developing lightweight, robust face recognition algorithms for resource-constrained mobile and embedded devices.
• Overcome limitations of unsupervised or handcrafted feature learning by introducing a fully supervised, hierarchical feature learning pipeline.
• Demonstrate that large-scale supervised training significantly improves recognition accuracy on benchmark datasets.
• Achieve high performance while maintaining low computational and memory overhead for real-world deployment.

Proposed method

• Model face images using receptive field theory, representing them as multi-channel Gaussian receptive maps, with only the most discriminative channels selected via supervised learning.
• Extract patches from the selected channels and perform over-complete patch pool search to identify the most discriminative patches using supervised criteria.
• Project each selected patch’s descriptor into a lower-dimensional subspace using discriminant subspace analysis (LDA) to enhance discriminability.
• Concatenate the learned features from all selected patches to form a complete face representation.
• Train a linear classifier on large-scale face pair data using ADMM (Alternating Direction Method of Multipliers) to scale training across a computing cluster.
• Apply the framework under strict protocols on FRGC and LFW benchmarks, using cross-validation and large-scale data to ensure robustness.

Experimental results

Research questions

• RQ1 Can a fully supervised hierarchical feature learning framework outperform existing handcrafted and unsupervised feature learning methods in face recognition?
• RQ2 Does large-scale training on a distributed cluster significantly improve recognition accuracy on standard benchmarks?
• RQ3 Can the proposed method maintain high accuracy while achieving low memory footprint and low computational cost for embedded deployment?
• RQ4 How does the hierarchical feature learning pipeline—spanning channel, patch, and descriptor levels—contribute to discriminative power?

Key findings

• The proposed method achieves a 92.6% true positive rate at 0.1% false positive rate on the FRGC-204 benchmark, outperforming all prior methods listed in Table 2.
• On the LFW benchmark under the image-restricted, label-free outside data protocol, the method achieves 91.54% accuracy with a standard error of ±0.49%, surpassing existing state-of-the-art approaches.
• The method is an order of magnitude faster in feature extraction than Gabor-based methods, which is critical for real-time applications.
• The system maintains a small memory footprint and low computational cost, making it suitable for deployment on embedded and mobile devices.
• Experiments confirm that increasing the number of training samples leads to notable accuracy improvements, validating the benefit of large-scale supervised learning.
• The hierarchical design—channel selection, patch selection, and LDA projection—proves effective in capturing discriminative features across multiple granularities.