A completed local shrinkage pattern for texture classification

• Published in: Applied soft computing 2020-12, Vol.97, p.106830, Article 106830
• Main Authors: Xu, Xiaochun, Li, Yibing, Wu, Q.M. Jonathan
• Format: Article
• Language: English
• Subjects: Completed local shrinkage pattern; Feature extraction; Texture classification; Texture image
• ISSN: 1568-4946; 1872-9681
• DOI: 10.1016/j.asoc.2020.106830

Visual texture classification plays a critical role in computer vision and pattern recognition. As one of the most popular texture descriptors, local binary pattern(LBP) has achieved extensive development and applications due to its simplicity and high efficiency. However, it is hard for most LBP-based methods to represent completed local texture information efficiently as they either only encode the sign of local difference or fuse low-discriminative features with high dimensionality. To alleviate these problems, this paper groundbreaking introduces a new insight in analyzing completed local texture information, called completed local shrinkage pattern(CLSP), which achieves a high tradeoff between discriminativeness and dimensionality. First, we map the completed local difference to a new encoding space through a shrinkage function and present the local shrinkage pattern. As the important supplement of local texture information, the center pixel is also encoded to build the local center pattern. Finally, the two sub-features are combined to generate the completed local shrinkage pattern. Moreover, we design a multi-information integration texture classification framework, which utilizes the original texture images and the gradient texture images to enrich the diversity of texture information. Experimental results on four popular texture databases demonstrate that the proposed CLSP descriptor achieves superior classification performance with low dimensionality. •A completed local shrinkage pattern is proposed to represent completed local texture information.•A multi-information integration texture classification framework is designed to enrich the diversity of texture information.•The proposed method has high classification accuracy.•The proposed method has low feature dimensions.