A ROBUST-TEXTURE CONVOLUTIONAL NEURAL NETWORK

Abstract

AlexNet was a breakthrough for the convolutional neural network (CNN) and showed the greatest successful mod- ified CNN that works well with large-scale images. However, it was unsuccessful in texture classification tasks. To extend CNN’s capability, this paper proposes a modified CNN architecture called a robust-texture convolutional neural network (RT-CNN) to serve both complex shape and texture classification tasks, especially in the following challenges: (i) the same class of images naturally contains various viewpoints, scales, uneven illuminations, etc. (ii) similarly shaped objects with different textures of images are often assigned into different classes; and (iii) dif- ferent shaped objects with similar textures of images are often assigned into the same class. The proposed scheme embeded a texture-embedded supplementary method, composed of texture compensation and supplement, into the CNN architecture. The texture compensation is constructed from texture subbands decomposed by 2D Littlewood-Paley empirical wavelet transform (2D Littlewood-Paley EWT). Then the texture supplement is constructed from texture subbands by using Gabor wavelet to extract multi-scale  and multi-orientation texture features.  Based  on two challenging datasets, the experimental results show that RT-CNN outperforms all test baseline methods: AlexNet, T-CNN, and wavelet-CNN, in terms of recognition accuracy rate. On a typical dataset, the recognition accuracy rate of the proposed method is still better than those of T-CNN and wavelet-CNN, and is comparable to that of AlexNet.