Depthwise Separable Convolutional Neural Network for Multi-class Texture Classification

This study proposed a compact deep learning-based texture classification model with Depth-wise Separable Convolution (DSC) layer. The method helps in extracting inherent texture features from the Convolutional Neural Network (CNN) architecture. The filters in depth-wise separable convolution layer act as filter bank which helps in classifying textures by minimizing the cross-entropy loss function. To speed up convergence and locate a better minimum for the loss function, an Adam optimization approach is utilized. The proposed model is experimented with the texture database: ALOT texture database, kth-tips2-b, kylberg texture dataset v1, kylberg sintorn rotation dataset. The results show that the suggested model achieve 97% accuracy on classifying texture images with lesser parameter.