Dual-branch network based on transformer for texture recognition

Yangqi Liu, Hao Dong, Guodong Wang, Chenglizhao Chen

Abstract: Texture recognition is the task of identifying and understanding the patterns and structures present in textures within digital images, constituting a highly challenging research topic. With the rapid advancement of deep learning, CNN-based methods have progressively supplanted traditional approaches, emerging as the predominant techniques for texture recognition. Nevertheless, these methods face limitations in capturing the structural information of textures from global features, attributed to the constraints of CNNs. Structural information stands out as a crucial factor in describing and recognizing real-world texture images. To address this challenge, we propose DBTrans, a dual-branch network based on Transformer architecture for texture recognition. DBTrans leverages the transformer's extraction capability for global features (low-frequency information) to capture the structural nuances of texture images. Additionally, recognizing the distinctive nature of texture recognition tasks, we propose a Transformer variant module tailored for this purpose, named Residual Pooling Transformer (RPT). Finally, we investigate the impact of different branches and modules on network performance and showcase the performance of DBTrans on five public datasets, achieving state-of-the-art accuracy.

External IDs:dblp:journals/dsp/LiuDWC24