PaperlayerView exact revision on arXiv

Source-linked AI summary

Discriminative Block-Diagonal Representation Learning for Image Recognition

Zheng Zhang, Yong Xu, Ling Shao, Jian Yang

arXiv:1707.03548v1cs.CVcs.LGcs.MM

TL;DR

Existing block-diagonal representation methods largely regularize training data, with less attention to jointly learning training and test representations. BDLRR combines off-block-diagonal suppression with a subspace model and low-rank criterion; across eight benchmark datasets and three recognition tasks, the paper reports results superior to state-of-the-art methods.

  • Problem

    Existing block-diagonal representation work mainly regularizes training data, with limited attention to jointly learning training and test representations.

    Source: §I · p. 2

  • Method

    BDLRR jointly optimizes a low-rank representation that suppresses off-block-diagonal coefficients and strengthens intra-class coherence through a subspace model.

    Source: §I · p. 2, §III · p. 4

  • Results

    Across eight public benchmark datasets and three recognition tasks, BDLRR is reported to outperform state-of-the-art methods.

    Source: §VI · p. 14

  • Takeaways & Limitations

    The reported results support BDLRR's potential for image recognition across the evaluated face, character, and scene-recognition datasets.

    Source: §VI · p. 14

    The semi-supervised model uses test data during learning, which may limit generalization when test data are unavailable at training time; the out-of-sample extension remains subject to algorithms' out-of-sample learning capability.

    Source: G. Limitation · p. 13

Abstract

from arXiv · show

Existing block-diagonal representation researches mainly focuses on casting block-diagonal regularization on training data, while only little attention is dedicated to concurrently learning both block-diagonal representations of training and test data. In this paper, we propose a discriminative block-diagonal low-rank representation (BDLRR) method for recognition. In particular, the elaborate BDLRR is formulated as a joint optimization problem of shrinking the unfavorable representation from off-block-diagonal elements and strengthening the compact block-diagonal representation under the semi-supervised framework of low-rank representation. To this end, we first impose penalty constraints on the negative representation to eliminate the correlation between different classes such that the incoherence criterion of the extra-class representation is boosted. Moreover, a constructed subspace model is developed to enhance the self-expressive power of training samples and further build the representation bridge between the training and test samples, such that the coherence of the learned intra-class representation is consistently heightened. Finally, the resulting optimization problem is solved elegantly by employing an alternative optimization strategy, and a simple recognition algorithm on the learned representation is utilized for final prediction. Extensive experimental results demonstrate that the proposed method achieves superb recognition results on four face image datasets, three character datasets, and the fifteen scene multi-categories dataset. It not only shows superior potential on image recognition but also outperforms state-of-the-art methods.

I. INTRODUCTION

The paper targets discriminative block-diagonal representations for image recognition, extending prior structured low-rank work by jointly learning training and test representations. BDLRR suppresses off-block-diagonal components and strengthens intra-class structure, with a subspace model and low-rank criterion supporting this joint learning.

I. INTRODUCTION

  • Prior sparse representations can overlook global structure, while structured low-rank approaches impose class structure but may not yield satisfactory discriminative features.The introduction identifies this gap as motivation for learning block-diagonal representations.Source: §I · pp. 1–2
  • The method uses a constructed subspace structure to improve training-sample self-expression and narrow the representation gap between training and test samples.A low-rank criterion is also used to capture correlations among classes or tasks.Source: §I · p. 2
  • BDLRR jointly learns training and test representations, suppressing off-block-diagonal components while strengthening block-diagonal structure.The semi-supervised framework is intended to improve extra-class incoherence and intra-class coherence.Source: §I · p. 2
  • The proposed optimization is solved with an ADMM-based alternating strategy, and the paper reports convergence analysis from theoretical and experimental perspectives.The introduction also notes an out-of-sample extension for new data instances.Source: §I · p. 2

A. Robust principal component analysis (RPCA)

The section reviews low-rank methods and develops BDLRR to learn discriminative block-diagonal representations for training and test samples. Its objective discourages extra-class coefficients while enhancing intra-class coherence and uses a low-rank criterion to capture class or task correlations.

A. Robust principal component analysis (RPCA)

B. Low-rank representation (LRR) based method

  • LRR addresses data drawn from multiple subspaces by representing observations with a low-rank coefficient matrix and an error term.Different error norms target different noise types, including Gaussian noise with the Frobenius norm and random noise or corruptions with the l1 norm.Source: B. Low-rank representation (LRR) based… · p. 3

III. THE PROPOSED BLOCK-DIAGONAL LOW-RANK REPRESENTATION

  • BDLRR jointly learns training and test representations by minimizing off-block-diagonal coefficients and enhancing coherent intra-class representations.Its subspace measure is relaxed to an elementwise l1 penalty because the l0 minimization is NP-hard; the low-rank criterion captures class or task correlations.Source: §III · pp. 3–4

IV. OPTIMIZATION AND ALGORITHM ANALYSIS

The optimization transforms the BDLRR objective into separable subproblems and updates variables iteratively with an alternating direction method. The resulting procedure uses closed-form updates and updates its Lagrange multipliers and penalty parameter during iterations.

IV. OPTIMIZATION AND ALGORITHM ANALYSIS

  • The BDLRR objective is solved by an alternating direction method that separates it into subproblems with closed-form solutions.An equivalent transformation introduces auxiliary variables to make the objective separable.Source: §IV · p. 4, A. Optimization Algorithm · p. 4

A. Optimization Algorithm

  • The augmented Lagrangian is minimized one coordinate direction at a time, fixing the remaining variables while updating each variable.The procedure updates the Lagrange multipliers and penalty parameter after optimizing the variables.Source: A. Optimization Algorithm · pp. 4–5
  • The recognition algorithm normalizes training and test samples, solves the BDLRR objective, learns a linear classifier, and predicts test labels.The learned representation contains separate training and test components.Source: A. Optimization Algorithm · p. 5

B. Recognition Method

The recognition procedure learns a linear classifier from the training representations and their labels, then uses it to predict labels for test samples.

B. Recognition Method

  • A linear classifier is learned from the training representation and its corresponding label matrix, then applied to test representations for prediction.The test sample's identity is determined using its representation in the learned test representation matrix.Source: B. Recognition Method · p. 6

C. Convergence Analysis

The proposed optimization is equivalent to a two-block ADMM, giving a convergence guarantee when the stated convexity and saddle-point conditions hold. The paper also reports empirical convergence for its chosen updates.

C. Convergence Analysis

  • Algorithm 1 is equivalent to a two-block ADMM, whose global convergence is theoretically guaranteed under the stated conditions.The formulation is cast in the classical ADMM form, and the theorem requires a saddle point plus specified convexity and rank conditions.Source: C. Convergence Analysis · p. 6
  • If a saddle point exists, the theorem states that the primal and dual sequences converge to it; otherwise, at least one specified sequence is unbounded.The theorem's conditional result concerns Z, W, and C converging to a saddle point, or γt or νt becoming unbounded.Source: C. Convergence Analysis · p. 6
  • The paper states that Algorithm 1 sets γt and νt to zero and reports that experimental convergence is preserved.The authors connect the optimization result to an optimal solution and report empirical convergence in Section V-F.Source: C. Convergence Analysis · p. 6

D. Computational Complexity Analysis

BDLRR's complexity includes an iteration-scaled cost dominated by matrix operations and SVD. Its burden is reported as slower than sparse-representation classifiers, slightly faster than regression methods, and generally comparable to low-rank/sparse methods.

D. Computational Complexity Analysis

  • The recognition process is described as linear in the sample number, whereas the SVD and matrix computations can become time-consuming for very large training and total sample counts.The passage specifies O(n^2N) for SVD of P when N > n.Source: D. Computational Complexity Analysis · pp. 6–7
  • BDLRR has total complexity Oκ(2n^2N + n^2d + dN + nN), with κ denoting the iteration count.The analysis identifies SVD and matrix computation as the main costs; computing SVD of P ∈ ℜ^(n×N) costs O(n^2N) when N > n.Source: D. Computational Complexity Analysis · p. 7
  • SRC, LRSI, and LatLRR are slower than BDLRR, while regression methods are a little faster and low-rank/sparse methods have generally comparable overall burden.The cited sparse methods solve (N − n) independent l1-norm minimization problems; the low-rank/sparse group also uses SVD and soft-thresholding.Source: D. Computational Complexity Analysis · p. 7

E. Out-of-sample Extension

The out-of-sample extension fixes the representation learned from available data and fits a representation for a novel image over the training samples. The resulting approach is presented as addressing recognition of instances beyond the original training and test data.

E. Out-of-sample Extension

  • BDLRR addresses the out-of-sample problem because recomputing the full model for unseen instances would be unrealistic and time-consuming.The preceding BDLRR stage learns representations only for the available samples.Source: E. Out-of-sample Extension · p. 7
  • For a novel image b, the extension learns representation z over Xtr while fixing the previously learned representation Z.It augments the earlier objective with terms for b while keeping the already learned variables.Source: E. Out-of-sample Extension · p. 7
  • Under the stated practical-data conditions, adding z does not change the nuclear norm because rank([Z, z]) equals rank(Z).The passage argues that z = Zα is underdetermined and, when n ≪ N, Z is row full rank.Source: E. Out-of-sample Extension · p. 7
  • The reduced representation-fitting problem is an elastic-net regularized regression problem, and the new instance is labeled by maximizing the learned classifier's response to z.The passage gives β1 = λ1/λ3 and β2 = λ2/2λ3; the recognition rule is label(b) = arg maxj(Ŵz).Source: E. Out-of-sample Extension · p. 7
  • The authors state that recognition of instances outside the training and test samples is well addressed by the proposed BDLRR model.They relate this result to learning discriminative block-diagonal training representations.Source: E. Out-of-sample Extension · p. 7

F. Discussion

The discussion characterizes BDLRR as combining supervised and semi-supervised learning properties with several representation-learning characteristics. It also situates the method alongside related discriminative low-rank approaches.

F. Discussion

  • BDLRR uses label information while learning training and test representations jointly in a semi-supervised formulation.Source: F. Discussion · p. 7
  • The method is described as inheriting properties of sparse, low-rank, structured, and elastic-net representation learning.Source: F. Discussion · p. 7
  • The discussion establishes relationships between BDLRR and NNLRS, SSLR, and SRRS.Source: F. Discussion · p. 8

1) Connection to the NNLRS method:

BDLRR jointly suppresses off-block representations and strengthens compact block-diagonal representations while learning training and test representations in a semi-supervised low-rank framework. It generalizes LRR and NNLRS, and is contrasted with SSLR’s representation constraints and separate learning stages.

1) Connection to the NNLRS method:

  • BDLRR generalizes LRR and NNLRS: both are special cases under specified parameter settings.Setting λ1=λ2=0 yields LRR; setting λ1=0 and X=Xtr yields weighted NNLRS without the nonnegative constraint.Source: 1) Connection to the NNLRS method · p. 8
  • NNLRS combines low-rank structure for global training-data relationships with sparse terms for each vector’s locality.BDLRR’s broader formulation also uses structured block-diagonal constraints for discriminative representations.Source: 1) Connection to the NNLRS method · p. 8
  • BDLRR jointly suppresses off-block-diagonal representations and strengthens compact block-diagonal representations while learning training and test representations in one semi-supervised framework.The method aims to enlarge margins between classes and enhance intra-class compactness.Source: 1) Connection to the NNLRS method · p. 8

2) Comparison with the SSLR method:

  • SSLR learns a structured dictionary using an ideal representation regularizer, then derives training and test representations in separate stages.Its representation objective uses a learned dictionary Ξ and a reconstruction term.Source: 2) Comparison with the SSLR method · p. 8
  • The authors question SSLR’s ideal-code constraint and note sensitivity to initialization, separate training/test learning, and multiple parameters to tune.They contrast this with BDLRR’s joint discriminative representation learning and training–test representation bridge.Source: 2) Comparison with the SSLR method · p. 8

3) Comparison with the SRRS method:

BDLRR is contrasted with SRRS, which learns a discriminative subspace from LRR-recovered clean data. In the reported face-recognition experiments, BDLRR outperformed the compared methods across four datasets and varied training-sample counts.

3) Comparison with the SRRS method:

  • SRRS applies discriminant analysis to LRR-recovered clean data and requires selecting the final representation dimension; BDLRR directly learns discriminative representations without that selection step.The authors also distinguish BDLRR’s joint learning of training and test representations from SRRS’s test representation obtained using P Xtt.Source: 3) Comparison with the SRRS method · p. 9

V. EXPERIMENTAL VALIDATION

  • The experiments assess recognition performance across datasets and also analyze algorithm convergence and parameter selection.The paper describes extensive comparisons with state-of-the-art recognition methods.Source: §V · p. 9

A. Experimental Setup

  • The evaluation covers eight benchmark datasets and three recognition tasks, with average accuracy reported over ten random training/test selections.For face datasets, parameter selection used five-fold cross-validation; standard splits were used for character datasets and the specified scene-recognition protocols were followed.Source: A. Experimental Setup · p. 9

B. Experiments for Face Recognition

C. Experiments for Character Recognition

BDLRR outperformed the compared methods on handwriting, scene-character, and fifteen-category scene recognition. The experiments cover three character datasets and the fifteen-category scene dataset, with scene-character recognition identified as a first application of discriminative representations.

C. Experiments for Character Recognition

  • The paper presents its work as the first to learn discriminative data representations for scene character recognition.The experiments use USPS handwriting and the Char74K and SVT scene-character datasets.Source: C. Experiments for Character Recognition · p. 11

1) Handwriting image recognition:

  • BDLRR performed consistently better than all compared methods across handwriting experiments using 30, 60, 90, or 120 training images per digit.The USPS dataset contains handwritten digits from 0 to 9.Source: 1) Handwriting image recognition · p. 11

2) Scene character image recognition:

D. Experiments for Scene Recognition

E. Experimental Analysis

BDLRR achieves the best recognition performance among compared methods across eight datasets, while experiments examine its convergence, parameter sensitivity, and learning-time constraints. The conclusion describes a semi-supervised representation model that jointly learns training and test representations for recognition.

E. Experimental Analysis

  • BDLRR gains the best recognition performance against compared state-of-the-art methods on eight datasets.The paper reports that transferring image features to BDLRR representations based on training features in a semi-supervised manner benefits recognition.Source: E. Experimental Analysis · p. 12
  • Although no method is best across all applications, BDLRR wins on low-resolution, limited-training experiments and shows a clearer test-set block-diagonal pattern than SRC.Figure 1 compares test-set representations from SRC and BDLRR on Extended YaleB, with representation values multiplied by 5.Source: E. Experimental Analysis · p. 12, Figure 1 · p. 13
  • BDLRR consistently outperforms CBDS and LatLRR on all datasets, globally minimizing off-block-diagonal components while compacting within-block structure.CBDS locally enforces class-wise diagonal structure; LatLRR extracts salient features, and the authors suggest lost image details may explain its weaker performance.Source: E. Experimental Analysis · pp. 12–13
  • BDLRR also performs well against recognition obstacles including face-image noise and scene-text complexity, low resolution, occlusion, blur, and text-size or font changes.The reported face-image conditions include occlusion, disguise, severe illumination changes, and expression variations.Source: E. Experimental Analysis · p. 13

1) Convergence study:

  • Across four datasets, relative error generally decreases with iterations and changes only slightly after 60 iterations, despite an early Extended YaleB fluctuation.The convergence curves cover Extended YaleB, AR, USPS, and Char74K-15.Source: 1) Convergence study · p. 13, Figure 2 · p. 14

2) Parameter Sensitiveness:

  • BDLRR is generally insensitive to λ1 and λ2; promising λ1 values avoid extremes, while λ2 performs best usually between 0.01 and 1.Performance is reported as not sensitive to λ3 in [10,25]; the authors suggest the simple Euclidean distance metric may explain the preference for smaller λ2.Source: 2) Parameter Sensitiveness · p. 13, Figure 3 · p. 14

G. Limitation

  • Because BDLRR jointly learns training and test representations, it requires test data during learning; otherwise, its results depend on out-of-sample methods.The authors note that training representations remain reasonably block-diagonal and report satisfactory recognition results in that setting.Source: G. Limitation · p. 13

VI. CONCLUSION

  • The conclusion presents BDLRR as a low-rank representation model that enhances extra-class incoherence and intra-class coherence, then uses an efficient linear classifier for recognition.The method collaboratively optimizes training and test representations and was evaluated on eight public benchmark datasets across three recognition tasks.Source: §VI · pp. 13–14

Related papers

Loading 1707.03548v1…