Texture Image Classification with Dilated Convolution Layers

Abstract

This work develops a compact deep-learning architecture to learn and recognize texture features. The suggested approach primarily concentrates on the feature-extracting layers of the neural network. The proposed Texture-Dilated Convolutional Neural Network (T-DCNN) is supported by blocks with dilated convolution layers. These blocks assist the model in retrieving the underlying texture attributes required for categorizing images. The built network was trained and evaluated on the kylberg texture database v.1.0. The model produced a result with 98.88% accuracy rate. The investigation shows that under same environment, the proposed model outperforms the conventional CNN model by lowering the required training time and parameters to categorize textures.