Source-linked AI summary
Structured Priors for Sparse-Representation-Based Hyperspectral Image Classification
Xiaoxia Sun, Qing Qu, Nasser M. Nasrabadi, Trac D. Tran
TL;DR
Sparse representation classification in HSI faces unstable coefficients, while structured priors exploit neighborhood relations and dictionary structure. This paper reviews and compares these priors and proposes a low-rank group prior. Its conclusion is that structured priors generally improve on classical ℓ1 sparsity, with performance depending on the prior and optimization method.
Problem
High dictionary mutual coherence can make HSI sparse coefficients nonunique or unstable, and structured priors address this by exploiting pixel dependencies or dictionary structure.
Source: §I · p. 1Method
The paper reviews and compares structured sparsity priors for SRC-based HSI classification and proposes the low-rank group prior.
Source: §I · p. 1, §IV · p. 5Results
Structured-prior SRC results generally improve over classical ℓ1 sparsity, while performance also depends on the prior and its optimization technique.
Source: §IV · p. 5Takeaways & Limitations
Low rank is more flexible than joint sparsity, which works better on large homogeneous regions, and dictionary group structure yields higher overall accuracy than ℓ1 sparsity.
Source: §IV · p. 5
Abstract
from arXiv · showhide
Pixel-wise classification, where each pixel is assigned to a predefined class, is one of the most important procedures in hyperspectral image (HSI) analysis. By representing a test pixel as a linear combination of a small subset of labeled pixels, a sparse representation classifier (SRC) gives rather plausible results compared with that of traditional classifiers such as the support vector machine (SVM). Recently, by incorporating additional structured sparsity priors, the second generation SRCs have appeared in the literature and are reported to further improve the performance of HSI. These priors are based on exploiting the spatial dependencies between the neighboring pixels, the inherent structure of the dictionary, or both. In this paper, we review and compare several structured priors for sparse-representation-based HSI classification. We also propose a new structured prior called the low rank group prior, which can be considered as a modification of the low rank prior. Furthermore, we will investigate how different structured priors improve the result for the HSI classification.
I. INTRODUCTION
Sparse representation classification (SRC) is affected by unstable sparse coefficients in HSI, prompting structured priors that use neighborhood dependencies, dictionary structure, or both. This paper compares such priors and proposes the low-rank group prior.
I. INTRODUCTION
- SRC represents a test pixel as a sparse combination of labeled training pixels and assigns its class using the minimum class-wise reconstruction residual.The dictionary concatenates class-wise sub-dictionaries.Source: §I · p. 1
- High dictionary mutual coherence can make HSI sparse coefficients nonunique or unstable, motivating priors that exploit neighboring-pixel dependencies or dictionary structure.The paper groups existing approaches into neighborhood-based, dictionary-based, and combined-structure priors.Source: §I · p. 1
- The paper assesses SRC with several structured sparsity priors for HSI classification and proposes the low-rank group prior, conceptually similar to CHiLasso.The priors reviewed include joint, Laplacian, group, sparse group, and low-rank sparsity.Source: §I · pp. 1–2
- The low-rank group prior assumes same-class pixels can be represented by sparse low-rank groups, combining group sparsity across groups with low-rank regularization within groups.The paper describes this as using one regularizer.Source: §I · p. 2
A. Joint Sparsity Prior
Joint sparsity models nearby HSI pixels as sharing the same sparse support, reflecting their typically similar materials and correlated spectra. It recovers a row-sparse coefficient matrix and classifies the center pixel by residual error.
A. Joint Sparsity Prior
- Joint sparsity assumes neighboring pixels share a common sparse support because nearby HSI pixels usually contain similar materials and have correlated spectra.For T neighboring pixels, their sparse coefficient vectors form the columns of X.Source: A. Joint Sparsity Prior · p. 2
- The shared support makes X row-sparse, and the model recovers this coefficient matrix through a Lasso problem.Only a few rows of X are nonzero.Source: A. Joint Sparsity Prior · p. 2
- The center pixel's class is selected using the minimum total residual error.The passage identifies this as the labeling step after joint-sparsity recovery.Source: A. Joint Sparsity Prior · p. 2
B. Laplacian Sparsity Prior
Laplacian sparsity uses similarities between neighboring pixels to encourage similar sparse coefficients without forcing every neighbor to share identical support. This flexibility accommodates neighborhoods that cross class boundaries.
B. Laplacian Sparsity Prior
- Joint sparsity can misclassify when correlated neighboring pixels recover partially different coefficients, while forcing identical atoms is unsuitable across class boundaries.Boundary neighborhoods may contain pixels from distinct classes that require different sub-dictionary atoms.Source: B. Laplacian Sparsity Prior · p. 2
- Laplacian sparsity weights pixel pairs by spectral similarity, encouraging similar pixels' coefficient vectors to be closer and dissimilar pixels' vectors to differ more.The weighting matrix is computed using sparse subspace clustering.Source: B. Laplacian Sparsity Prior · p. 2
- Unlike joint sparsity, the Laplacian prior does not require all neighboring pixels to have the same support, and a modified feature-sign search algorithm can solve its optimization.Its grouping constraint is applied to testing pixels rather than dictionary atoms.Source: B. Laplacian Sparsity Prior · p. 2
C. Group Sparsity Prior
Group sparsity exploits the class-wise sub-dictionaries in SRC by selecting groups of atoms rather than individual atoms. It can be extended to neighboring pixels, and adding an ℓ1 penalty encourages sparsity within selected groups.
C. Group Sparsity Prior
- Because the SRC dictionary consists of class-wise groups, group Lasso encourages coefficients from selected groups to be active while leaving other groups inactive.The passage notes its relevance when input pixels are inherently mixed.Source: C. Group Sparsity Prior · p. 2
- Group Lasso extends to multiple neighboring pixels as collaborative group Lasso, and a group size of one reduces it to joint sparsity Lasso.The collaborative formulation uses group coefficient matrices.Source: C. Group Sparsity Prior · p. 3
D. Sparse Group Sparsity Prior
The sparse group prior adds within-group sparsity to group Lasso when a sub-dictionary is overcomplete, and the formulation is extended to multiple features.
D. Sparse Group Sparsity Prior
- When a sub-dictionary is overcomplete, an ℓ1-norm term is added to group Lasso to enforce sparsity within each group.The passage introduces this as an extension of the group Lasso formulation.Source: D. Sparse Group Sparsity Prior · p. 3
- The sparse group Lasso formulation is extended to the multiple-feature case.Source: D. Sparse Group Sparsity Prior · p. 3
- The single-feature problem is called sparse group Lasso, while its multiple-feature form is called collaborative hierarchical Lasso (CHiLasso).Source: D. Sparse Group Sparsity Prior · p. 3
E. Low Rank/Group Sparsity Prior
The low-rank prior is extended to account for dictionary groups. The resulting low-rank group prior combines sparsity within and across groups using one regularization term.
E. Low Rank/Group Sparsity Prior
- The low-rank prior is more flexible than joint sparsity, and is reported to outperform it when neighboring pixels occupy small non-homogeneous regions.Source: E. Low Rank/Group Sparsity Prior · p. 3
- The low-rank group prior extends the low-rank prior to dictionary groups by summing the ranks of group coefficient matrices.Source: E. Low Rank/Group Sparsity Prior · p. 3
- The toy-example figure displays sparsity patterns for low-rank sparsity and low-rank group sparsity alongside other priors.Source: Figure 1 · p. 3
- Minimizing each group’s nuclear norm yields within-group sparsity, while summing those norms yields group sparsity across groups.The prior achieves both patterns with a single regularization term.Source: E. Low Rank/Group Sparsity Prior · p. 3
III. RESULTS AND DISCUSSION
The experiments assess structured sparsity priors on two hyperspectral images and a toy example. They describe the datasets, evaluation measures, and implementations used for comparison.
A. Datasets
- The evaluation uses Indian Pine and University of Pavia images plus a toy example; 997 labeled pixels are selected for training in each image, with the rest used for testing.Indian Pine has 16 ground-truth classes; University of Pavia has nine.Source: A. Datasets · p. 3
B. Models and Methods
- The tested priors are joint, Laplacian, collaborative group, sparse group, low-rank, and low-rank group sparsity.Source: B. Models and Methods · pp. 3–4
- The supplied visuals cover Indian Pine classification maps and accuracy, University of Pavia accuracy, and Indian Pine computation time.Source: Figure 2 · p. 4, Table II · p. 4, Table IV · p. 4
- Image classification is evaluated on the test set using overall accuracy (OA), average accuracy (AA), and the κ coefficient; each prior is reported at its cross-validated highest OA.Source: B. Models and Methods · p. 4
- The comparison uses ADMM for joint, group, and low-rank priors, SpaRSA with ADMM for CHiLasso and Laplacian, and modified FSS for Laplacian Lasso.The authors organize the implementations into ADMM/SpaRSA and FSS-based categories.Source: B. Models and Methods · p. 4
C. Results
Results vary with image structure and optimization: Laplacian sparsity via FFS gives the highest OA on both reported images, while the priors show distinct sparsity patterns in a toy example.
C. Results
- In the toy example, joint sparsity activates many incorrect rows, while SGS deactivates more atoms within groups than GS and LRG resembles SGS.Laplacian sparsity also shows similarity among coefficients belonging to the same classes.Source: C. Results · p. 5
- On Indian Pine, Laplacian sparsity via FFS gives the highest OA, while both SGS and LRG outperform GS.Joint sparsity performs better than low rank on this image’s large homogeneous regions; among ADMM-based methods, LRG produces the smoothest result.Source: C. Results · p. 5
- On University of Pavia, Laplacian sparsity via FFS gives the highest OA, but differences among structured priors are small.Among ADMM-optimized priors, group sparsity gives the highest OA; low rank outperforms joint sparsity in this image’s many small homogeneous regions.Source: C. Results · p. 5
IV. CONCLUSION
The paper reviews five structured sparse priors, proposes low-rank group sparsity, and concludes that results depend on prior structure and optimization technique.
IV. CONCLUSION
- The paper reviews five structured sparse priors and proposes a low-rank group sparsity prior.Source: §IV · p. 5
- Structured-prior SRC classification results generally improve over classical ℓ1 sparsity, and imposing a group-structured prior on the dictionary gives higher OA than ℓ1.Source: §IV · p. 5
- The low-rank prior is more flexible than joint sparsity, while joint sparsity works better on large homogeneous regions.Source: §IV · p. 5
- Performance depends not only on the structured prior but also on its corresponding optimization technique.Source: §IV · p. 5