Dynamic Structured Discriminant Dictionary Pair Learning for Pattern Classification

doi:10.1109/access.2026.3718299

Dynamic Structured Discriminant Dictionary Pair Learning for Pattern Classification

2026-01-01

0

PRE

AI

张永良 (Yonghao Zhang) *

DOI：10.1109/access.2026.3718299

Original

Original request for help

Save

Abstract

En 中文

Discriminative analysis-synthesis dictionary pair learning is effective for pattern classification. However, most methods use the graph preserving criterion to transmit the local structure relationships of samples, which weaken the performance and generalization capability of model. To precisely capture local neighborhood relationships of samples and enhance the discrimination capability of analysis-synthesis dictionary pair, we propose a dynamic structured discriminant dictionary pair learning (DSDDPL) method. First, we design a dynamic local structure transmission term that uses a transmission matrix to adaptively transmit the local structure information of samples to the coding coefficients to improve the discrimination capability of the coding coefficients. Second, we design a discriminative sparse code error term to constrain the coding of training samples, guiding the coefficient matrix to an ideal block-diagonal structure to improve the discrimination capability of dictionary pair indirectly. Finally, these two terms integrate into the basic discriminative dictionary pair learning model to form DSDDPL. Moreover, we design an efficient iterative algorithm to solve the optimization problem of DSDDPL. We construct the comparison experiment to evaluate DSDDPL with nine discriminative dictionary learning methods on six datasets. DSDDPL achieves the highest classification accuracy on the five datasets and the second highest on CLD 22 dataset. The experimental results show that DSDDPL achieves superior classification performance compared to existing dictionary learning methods. DSDDPL utilizes the local structure information and label information of samples to constrain the coding coefficients, which further improves the discrimination capability of the analysis dictionary and the synthesis dictionary. DSDDPL provides an effective technical solution for pattern classification in small-scale datasets, including facial recognition and crop disease detection.

Keywords:

Dictionaries

Licenses

Machine learning

Modeling

Nuclear facility regulation

Accuracy

Codes

Matrices

Training

Weighted sum model

Dictionary learning

dictionary pair learning

sparse codes

pattern classification

Dynamic Structured Discriminant Dictionary Pair Learning for Pattern Classification

Abstract

Journal

Organization

Cited Papers