TSDAnet: texture strip dual attention network for intraclass texture classification

Intraclass texture classification is a challenging task due to the high similarity between subtypes within the same texture. The presence of comparable texture properties degrades the performance of convolutional neural networks by misclassification. In this work, a texture strip dual attention network is developed to appropriately understand the relevant characteristics of similar textures. The novel Bidirectional Texture Strip Attention (BTSA) module in TSDAnet generates dual attention features from discrete channels and bidirectional strips using long-narrow kernels. The fused features in the BTSA module harvest the contextual information for each pixel under the guidance of the learned weights via depthwise convolutional filters. In addition, a mixed pooling layer is incorporated into the proposed network with an adaptive selection pooling mechanism. This layer aggregates local features by introducing randomness in the choice of average or maximum pooling on a specified window. The efficiency of the proposed network is studied in detail on three texture datasets (GTOS-mobile, KTH-TIPS2b, and STex) with seven colour models. The results reveal that TSDAnet can classify intraclass textures with 99.3% accuracy in the HSI colour model.