Multi scale supervised entropy weighted binary pattern for texture classification

Texture is a crucial visual and sensory attribute in understanding the world. The complexity of imaging environments, variations in acquisition angles and distances, and differences in resolution make representing multi-scale texture features a core challenge in texture analysis. However, most existing multi-scale methods are overly complex and redundant, often neglecting the correlation of texture features across different scales. To tackle these challenges, this paper proposes an efficient multi-scale supervised entropy-weighted binary pattern for texture classification. Firstly, this paper introduces a local entropy-weighted histogram based on two-dimensional entropy to enhance the discriminative power of binary pattern operators. Secondly, to select the optimal texture scale from the Gaussian scale space, the paper proposes a local entropy-based optimal selection mechanism (LEOSM) grounded in the uniform properties of the proposed local entropy-weighted histogram. A local entropy-based optimal selection mechanism (LEOSM) is designed to adaptively select representative texture scales from the Gaussian scale space, based on the uniformity properties of the proposed local entropy-weighted histogram, thereby enhancing scale robustness. Thirdly, a cross-scale uniformity supervised pattern framework (CSUSPF) is proposed to jointly encode multi-scale and cross-scale texture information, enabling a more compact, abstract, and discriminative representation. In addition, a novel cross-scale entropy-weighted center pattern and an entropy-weighted entropy pattern are proposed to capture interrelationships among texture images in the Gaussian scale space, thereby mitigating potential information loss across scales. To validate the effectiveness of the proposed method for texture classification tasks, a series of experiments were conducted on five public texture datasets: Outex, UIUC, CUReT, UMD and ALOT. These results indicate that the proposed method consistently outperforms the baseline CLBP by a margin of 1-5%, and also achieves superior performance compared to several state-of-the-art approaches.