Edge Structure Learning via Low Rank Residuals for Robust Image Classification

Xiangjun Shen, Stanley Ebhohimhen Abhadiomhen, Yang Yang, Zhifeng Liu, Sirui TianAAAI 2023 · DOI

Abstract

High-dimensional image data is frequently compromised by redundancy and corruption, such as noise or occlusions. While Low-Rank Representation (LRR) is a standard technique for data recovery, it often filters out critical high-frequency edge details as "residuals." This paper presents ESL-LRR (Edge Structure Learning via Low Rank Residuals), a novel framework that treats these residuals as a secondary informative view rather than mere noise. By preserving discriminative edge structures, ESL-LRR significantly improves classification robustness across various benchmark datasets.

Introduction

In computer vision, distinguishing between objects often relies on sharp edges and fine details. However, traditional low-rank methods prioritize global structures (low-frequency signals), effectively acting as low-pass filters. When images are heavily corrupted, the loss of these edges makes classification tasks—such as facial recognition or digit identification—extremely difficult. The core challenge addressed by this research is the effective separation of useful edge information from useless corruption within the residual matrix.

Methodology: The ESL-LRR Framework

The proposed ESL-LRR method shifts the paradigm by utilizing the residual matrix ($E$) as a valuable data source. The framework integrates several key components:

  1. Manifold Learning on Residuals: A dynamic affinity graph is constructed to capture the geometric structure of residuals, identifying consistent edge features across samples.
  2. Adaptive Projection: The model learns a projection matrix ($P$) specifically from the residual space. The dynamic graph allows the model to adaptively weight consistent structures while suppressing random noise.
  3. Coupled Regularization: The objective function simultaneously seeks an intrinsic low-rank representation and discriminative, edge-preserved projections.
Decomposition of images into low-rank and residual components. The residual captures the vital edge structures often lost in traditional LRR.

The optimization of this complex objective is handled via an efficient Augmented Lagrange Multiplier (ALM) strategy, ensuring iterative convergence.

Experimental Evaluation

The researchers evaluated ESL-LRR on five benchmark datasets: MNIST, USPS, Yale, LFW, and COIL100. To test robustness, datasets were subjected to pepper noise (up to 40%) and block occlusions.

Key Results

  • Robustness to Noise: On the MNIST dataset with 40% noise, ESL-LRR achieved 91.41% accuracy. On the Yale Faces dataset, it maintained 78.02% accuracy under the same noise level, significantly outperforming traditional LRR-based methods.
  • Occlusion Handling: In tests using the LFW dataset with 18x18 block occlusions, ESL-LRR reached 75.08% accuracy, a 5% improvement over the competitive DLRRPD method.
  • Performance Gain: Across all datasets, ESL-LRR consistently outperformed state-of-the-art methods like Low-Rank Embedding (LRE) and Feature Selective Projection (FSP) by over 2% on average.

Discussion and Conclusion

While ESL-LRR provides significant accuracy gains, it faces challenges regarding parameter sensitivity (requiring tuning of three hyperparameters) and computational intensity due to its multi-block optimization. However, the fundamental insight—that residuals are a treasure trove of edge information—paves the way for more robust computer vision systems. By combining global low-rank structures with local edge features, the framework achieves superior performance in challenging, corrupted environments. Future work will focus on scaling the optimization process and integrating edge-learning logic into deep learning architectures.

This summary is based on the paper "Edge Structure Learning via Low Rank Residuals for Robust Image Classification" presented at the Thirty-Seventh AAAI Conference on Artificial Intelligence (AAAI-23).

Summary updated

BibTeX
@inproceedings{shen2023edge,
  author = {Xiangjun Shen and Stanley Ebhohimhen Abhadiomhen and Yang Yang and Zhifeng Liu and Sirui Tian},
  title = {{Edge Structure Learning via Low Rank Residuals for Robust Image Classification}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.1609/aaai.v37i2.25318},
}
Figure