An Accessory Detection Algorithm Based on Sparse Representation and Low-Rank Matrices

Abstract

To address the issues of low detection accuracy, poor robustness, and limited generalization in accessory detection under complex scenes, we propose an accessory detection algorithm that integrates sparse representation with low-rank matrices. First, discriminative low-rank matrix recovery is employed to correct poor-quality training samples. By learning a low-rank projection matrix, the feature matrix of the test sample is projected onto a corresponding low-rank subspace. These two matrices are then used to construct a redundant dictionary for sparse representation. Classification is performed using the sparse representation approach, and the sparse representation coefficients are obtained by solving the problem with the Adaptive Penalty Linearized Alternating Direction Method (LADMAP). Finally, the reconstruction error for each class is calculated using the sparse representation coefficients, thereby achieving accurate detection of accessories worn by people. The results show that the proposed method achieves an average accuracy of 83.71% with 38.78 million parameters. Compared to the baseline model YOLOv5s, the number of parameters is reduced by 40.88%, while the average accuracy is improved by 16% and the precision is increased by 0.1507, demonstrating excellent performance in accessory detection.