Abstract: The miscellany of fabric and the complexity of its texture have been viewed as enormous challenges to distinguish artificially.The merging multi-scale attention co-occurrence representation's residual texture encoding network(MACTEN) has been proposed with the introduction of the deep learning technology.And based on that,the large-scale fabric classification system on the web has been carried out.The MACTEN mainly composed of attention co-occurrence representation module (ACM) and improved residual coding module (REM),as well as multi-scale texture coding fusion module (MTEM).In this work,the mechanism of attention has been implemented into ACM to deal with different types of clothes,which adaptively adjusts the weight of texture co-occurrence features,and optimizes the joint distribution of co-occurrence features by expanding the co-occurrence domain to form more refined texture co-occurrence features.Moreover,the improved residual coding,including global texture information of spatial invariance,has been obtained with introduction of dictionary learning method into REM,which can solve the problem of disordered representation of cloth texture effectively.Finally,MTEM combined multiple scale attention texture co-occurrence features and cascaded residual texture coding as descriptors,can represent different shape and size of disordered fabric texture.On self-building cloth dataset,MACTEN has exhibited better performance than other baseline algorithms.Furthermore,the experimental results of KTHTIPS,FMD and DTD datasets show that MACTEN can be generalized as a general texture classification algorithm.

Key words: Cloth classification, Co-occurrence feature, Deep learning, Feature fusion, Texture descriptors