Hierarchical heterogeneous ant colony optimization based weight generation for texture classification

Abstract

Texture classification is an important problem in pattern recognition and computer vision. The goal of texture analysis is to represent texture in a model that is invariant to changes influenced by illumination, rotation and noise. Natural images exhibit structures that are highly complex and therefore texture analysis have turned a challenging problem. Moreover, textures in real-time environment contain textural information at varying scales. Local Binary Pattern (LBP) is an effective non-parametric texture operator that encodes the local structure around each pixel. This paper proposes a Similarity based Texture Classification for LBP (STC-LBP) algorithm for classification of texture images. STC-LBP algorithm classifies the query texture image based on the similarity of LBP features. To find an optimal weight that emphasizes the similarity between features, a Hierarchical Heterogeneous Ant Colony Optimization based Weight Generation (HHACOWG) algorithm is proposed. Experiments were performed on five benchmark texture datasets namely KTH-TIPS, Brodatz, CUReT, Outex_TC10 and Outex_TC12 datasets. Experiments reveal that the proposal achieves better classification accuracy when compared to the state-of-art methods. The scale and rotation invariance property of STC-LBP is significantly better compared to the existing texture descriptors. The proposal is also tolerant to noise and illumination changes. The results of experiments were validated using non-parametric statistical tests. The feature dimension of the proposal is significantly less compared to the existing descriptors for texture classification. The time taken for feature extraction is less compared to the existing methods indicating that the proposal is well suited for real-time applications.