Two-dimensional PLSR with manifold optimization based CNN for image classification

doi:10.1007/s13042-025-02593-1

Two-dimensional PLSR with manifold optimization based CNN for image classification

2025-03-01

0

PRE

AI

DOI：10.1007/s13042-025-02593-1

Original

Original request for help

Save

Abstract

En 中文

Traditional Partial Least Squares Regression (PLSR) aims to establish a linear relationship between independent and dependent variables. It will convert the original data into one-dimensional vectors, thus destroying the structural information of the data and generating dimensional catastrophes. To address these limitations, this paper proposes a novel model CNNM2DPLSR called Two-Dimensional Partial Least Squares Regression (2DPLSR) with Manifold Optimization based Convolutional Neural Network (CNN). Firstly, deep features are extracted from the original data using CNN, which are taken as the new independent variables. To retain the structural information of the data, after performing bilateral dimensionality reduction on the independent variables, the resulting low-dimensional features are multiplied by the dependent variable to maximize the inner product, thereby obtaining the bilateral dimensionality reduction matrix. Finally, an objective function for the model is constructed and undergoes a detailed derivation process. Given that the dimensionality reduction matrix satisfies the column orthogonality constraints, the manifold optimization approach is operated in this model to obtain more accurate numerical solutions. Experimental results on various datasets show that the proposed method has lower classification error rates and better adaptability than other representative methods.

Keywords:

2DPLS

Feature extraction

Dimensionality reduction

Manifold optimization

CNN

Two-dimensional PLSR with manifold optimization based CNN for image classification

Abstract

Journal

Organization