TexFusionNet: An Ensemble of Deep CNN Feature for Texture Classification

Abstract

The texture classification from images is one of the important problems in pattern recognition. Several hand-designed methods have been proposed in last few decades for this problem. Nowadays, it is observed that the convolutional neural networks (CNN) perform extremely well for the classification task mainly over object and scene images. This improved performance of CNN is caused by the availability of huge amount of images in object and scene databases such as ImageNet. Still, the focus of CNN in texture classification is limited due to non-availability of large texture image data sets. Thus, the trained CNN over Imagenet database is used for texture classification by fine-tuning the last few layers of the network. In this paper, a fused CNN (TexFusionNet) is proposed for texture classification by fusing the last representation layer of widely adapted AlexNet and VGG16. On the top of the fused layer, a fully connected layer is used to generate the class score. The categorical cross-entropy loss is used to generate the error during training, which is used to train the added layer after the fusion layer. The results are computed over several well-known Brodatz, CUReT, and KTH-TIPS texture data sets and compared with the state-of-the-art texture classification methods. The experimental results confirm outstanding performance of the proposed TexFusionNet architecture for texture classification.