Learning a minimum similarity projection and lowest correlation representation for image classification. Issue 4 (19th February 2020)

• Record Type: Journal Article
• Authors: Lu, Shan; Zhang, Jun; Gao, Ying
• Is Part Of: IET image processing. Volume 14:Issue 4(2020)
• Page Start: 782
• Page End: 788
• Publication Date: 2020-02-19
• Subjects: feature extraction -- face recognition -- image representation -- learning (artificial intelligence) -- image classification -- matrix algebra; minimum similarity projection -- lowest correlation representation -- representation learning -- image classification problem -- classification performance -- image classification method -- discriminative representation -- low‐dimensional space -- feature projection -- feature representation -- projection matrix -- representation results -- learned projection -- visual classification tasks -- scene classification; Image processing -- Periodicals; 621.36705
• DOI: 10.1049/iet-ipr.2019.0683
• Languages: English
• ISSNs: 1751-9659

Abstract

Abstract : The projection and representation learning is an attractive tool for image classification problem due to its effectiveness and efficiency of extracting interior structure for data. However, the complexity and diversity of real data lead to the decline of classification performance. A novel image classification method is proposed by learning a minimum similarity projection and lowest correlation representation. This method attempts to produce a discriminative representation on a low‐dimensional space for the data, which takes two steps: feature projection and feature representation. By learning a projection matrix, the feature projection aims to map the samples into a low‐dimensional space which jointly minimises the similar within‐class difference and maximises the dissimilar cross‐class difference. A discriminative representation for the data on the new space is generated by using the de‐correlated effect to the representation results of all classes. Therefore, the learned projection and representation simultaneously demonstrate discriminative properties in the learning of both steps. The extensive experiments conducted on different visual classification tasks consist of face recognition, object categorisation, and scene classification that the proposed method performs superior performance for image classification.