Source-linked AI summary
Collaborative Representation based Classification for Face Recognition
Lei Zhang, Meng Yang, Xiangchu Feng, Yi Ma, David Zhang
TL;DR
SRC’s working mechanism and its use of training samples from all classes to represent a query are not fully revealed or sufficiently examined. The paper presents collaborative representation as a general model, reports competitive accuracy with lower time complexity, and identifies conditions where l2-norm residual modeling can outperform l1-norm modeling.
Problem
SRC’s working mechanism is not fully revealed, while its use of training samples from all classes to represent the query sample is rather ignored.
Method
The paper presents collaborative representation as a more general model, with different settings of p and q producing different instantiations.
Results
The l2-regularized CRC-RLS and R-CRC have very competitive face-recognition accuracy but much lower time complexity than regularized SRC.
Takeaways & Limitations
Collaborative representation plays a more important role than l1-regularization of the coding vector in face representation and recognition.
Takeaways & Limitations
The residual errors in query images are not always truly sparse or approximately Laplacian; with real scarves, l2-norm modeling can work better than l1-norm modeling.
Abstract
from arXiv · showhide
By coding a query sample as a sparse linear combination of all training samples and then classifying it by evaluating which class leads to the minimal coding residual, sparse representation based classification (SRC) leads to interesting results for robust face recognition. It is widely believed that the l1- norm sparsity constraint on coding coefficients plays a key role in the success of SRC, while its use of all training samples to collaboratively represent the query sample is rather ignored. In this paper we discuss how SRC works, and show that the collaborative representation mechanism used in SRC is much more crucial to its success of face classification. The SRC is a special case of collaborative representation based classification (CRC), which has various instantiations by applying different norms to the coding residual and coding coefficient. More specifically, the l1 or l2 norm characterization of coding residual is related to the robustness of CRC to outlier facial pixels, while the l1 or l2 norm characterization of coding coefficient is related to the degree of discrimination of facial features. Extensive experiments were conducted to verify the face recognition accuracy and efficiency of CRC with different instantiations.
1. Introduction
The paper argues that SRC’s collaborative use of all training samples matters more to face-recognition success than l1 sparsity alone, motivating the broader CRC framework.
- SRC mechanism: SRC codes a query face over training samples from all classes and classifies it using class-specific reconstruction residuals.For a query from class i, coefficients associated with Xi are expected to dominate.
- Motivation: SRC’s l1-minimization can be time consuming, prompting algorithms that improve its computational speed.The paper cites Gradient Projection, Homotopy, Iterative Shrinkage-Thresholding, Proximal Gradient, and ALM approaches.
- Central claim: The paper identifies collaborative representation as more crucial to face recognition than l1 sparsity and reports similar performance from cheaper l2 regularization.This reframes SRC as relying substantially on representing the query collaboratively across classes.
- Collaborative representation: Using all classes to represent a query alleviates the small-sample-size problem, especially when each class has few training samples.SRC differs from nearest classifiers that represent the query separately within each class.
- CRC framework: CRC generalizes SRC by varying norms on the coding residual and coefficient vector, linking residual norms to outlier robustness and coefficient norms to feature discrimination.S-SRC and R-SRC are presented as special cases of CRC.
- Coefficient discrimination: For highly discriminative features, either l1- or l2-regularized coefficients can concentrate on same-class samples; l1 helps when feature dimensionality is too low.The paper relates coefficient sparsity and discrimination to the dimensionality of the facial feature.
2. The Role of Sparsity in Representation based Face Recognition
This section examines whether sparsity itself drives discrimination, showing that regularization and collaborative representation can distinguish classes even when l2 coding is non-sparse.
- SRC design: SRC combines sparse coefficient coding with representation over the whole training dataset rather than coding separately within each class.These are identified as two key points of standard SRC.
- Collaborative representation: Collaborative representation helps address limited samples by offering more training atoms for representing the query.The discussion contrasts SRC with class-wise nearest-subspace representations.
- Class similarity: When class dictionaries are highly similar, small residual differences can make nearest-subspace classification unstable under noise or deformation.A small dictionary difference Δ produces similarly small residual differences, allowing the wrong class to win.
- Regularization: Using fewer samples to achieve comparable reconstruction accuracy can improve class discrimination, motivating l0, l1, and l2 coefficient regularizers.The example contrasts roughly 5 or 6 samples for the correct class with 8 or 9 for a similar incorrect class.
- Empirical illustration: With more than 10 samples, residual discrimination can decrease, whereas stronger l1 or l2 regularization can restore discriminative reconstruction.The illustrated curves use a query from Extended Yale B class 32 and compare it with similar class 5.
- Empirical illustration: Smaller ε makes both l1 and l2 coefficient regularization produce discriminative residuals, and l2 can match sparse regularization in this example.Weak regularization at ε > 8 yields low residuals for both classes and unstable classification.
- Practical implications: The paper notes that l0 minimization is combinatorial and NP-hard, making l1 or l2 regularization more practical.It also describes regularized nearest subspace as an alternative when each class has many training samples.
3. Collaborative Representation based Classification (CRC)
CRC classifies a query by collaboratively representing it with samples from all classes, then evaluating each class’s contribution and reconstruction residual. The paper argues that this collaborative mechanism is more important than l1 sparsity, while l2 regularization can provide comparable recognition with lower complexity.
- Motivation: Face similarities allow samples from other classes to help represent a query, but selecting such nonlocal samples directly is difficult and computationally expensive.Distance-based selection may be ineffective for classification, and adding nonlocal samples can reduce residual discrimination.
- Collaborative representation: CRC addresses limited per-class training data by representing each query over the samples from all classes.This avoids separately searching for nonlocal samples and gives all classes a shared representation.
- Classification rule: CRC assigns a query to the class contributing most to its collaborative representation, equivalently the class with the least reconstruction residual.Its geometric interpretation double-checks both alignment with the query and separation from contributions of other classes.
- Efficiency and robustness: l2 regularization achieves similar face-recognition results to l1 regularization with significantly lower complexity, supporting the greater importance of collaboration over l1 sparsity.The CRC-RLS decision combines reconstruction residual and coefficient sparsity, slightly improving accuracy over residual-only classification.
- Coefficient sparsity: Collaborative representation can make coding coefficients naturally sparse, especially with discriminative high-dimensional features, reducing the need for expensive l1 regularization.Lower feature dimensionality weakens discrimination and makes the coefficients less sparse.
4. Experimental Results
Experiments investigate whether collaborative representation or l1 sparsity drives face-recognition performance, comparing CRC instantiations across accuracy, robustness, and efficiency settings. Results consistently show strong performance from collaborative methods, with CRC-RLS often matching or exceeding sparse alternatives while substantially reducing runtime.
- Experimental aim: The experiments test CRC variants to separate the roles of sparsity and collaborative representation rather than simply claim CRC is more accurate than SRC.SRC is treated as a CRC instantiation using l1-norm regularization, while experiments compare coding norms and collaborative mechanisms.
- Parameter sensitivity: Small positive λ values produce good S-SRC and CRC-RLS results, whereas λ=0 fails and λ>0.1 reduces recognition rates.The reported useful range begins around 0.000001 and extends to 0.1.
- Sparsity and discrimination: When facial features are sufficiently discriminative, CRC-RLS achieves recognition rates similar to S-SRC across a broad λ range, showing l1 sparsity is not necessary.The paper attributes this result to the discrimination of the facial feature rather than to a need for sparse coefficients.
- Training-sample regime: With sufficient training samples per class, RNS_l1 and RNS_l2 obtain the same best results, while CRC-RLS is about 1.4% higher than S-These comparisons indicate that coding each class can be more effective than coding all classes when per-class training data are sufficient.
- Recognition accuracy: CRC-RLS and S-SRC generally lead competing methods across face-recognition datasets, including LFW, where CRC-RLS performs slightly better than S-SRC under uncontrolled variation.On LFW, all methods perform worse than on several controlled databases because of pose and misalignment variation.
- Robustness: For occlusion and corruption, R-CRC improves over R-SRC by about 13% at 80% corruption and remains effective up to 50% pixel corruption.The robust l1-residual formulation is reported as more robust than l2 regularization, while CRC-RLS can still perform well in several disguise settings.
- Efficiency: CRC-RLS combines near-SRC accuracy with large speed advantages: it is 1636 and 38.8 times faster than the compared S-SRC implementations on Extended Yale B.On Multi-PIE it is more than 1600 times faster with only a 0.4% lower recognition rate, and on a large-scale database it is about 40 times faster than Homotopy while improving recognition by more than 2%.
5. Conclusions and Discussions
The paper argues that collaborative representation is more important to face-recognition performance than l1 sparsity, and presents CRC variants that retain competitive accuracy with lower complexity. It also relates norm choices to feature discrimination and robustness to facial outliers, while identifying settings where l1 regularization or residual modeling has limitations.
- Collaborative representation plays a more important role in face representation and recognition than l1 regularization of the coding vector.
- CRC generalizes SRC, with CRC-RLS for unoccluded faces and robust CRC for occlusion or corruption.
- CRC-RLS and R-CRC achieve face-recognition accuracy competitive with l1-regularized SRC while requiring much lower time complexity.
- The benefit of l1 regularization depends on feature dimension: high-dimensional features can be naturally sparse, whereas low-dimensional features may benefit from active sparsification.
- l1-regularized coding can force an invalid query into one class, whereas l2-regularized coding distributes it across classes and avoids this problem.
- l1 residual characterization can tolerate sparse or Laplacian-like outlier pixels, but its effectiveness is limited when query-image errors do not follow that structure.
- For real scarf occlusion, errors are not sparse and l2-norm residual modeling can outperform l1-norm.