Robust Deep Gaussian Descriptor for Texture Recognition

Recently, second-order statistical modeling methods with convolutional features have shown impressive potential as image representation for vision tasks. Among them, bilinear convolutional neural network (B-CNN) has attracted a lot of attentions due to its simplicity and effectiveness. It captures the second-order local feature statistics via outer product, which approximately explores the covariance between convolutional features and achieves promising performance for texture recognition. In order to inherit the merits of B-CNN while further improving its performance, we introduce a Gaussian descriptor into B-CNN and propose a novel robust deep Gaussian descriptor (RDGD) method for texture recognition. We first compute Gaussian by using the output of outer product of B-CNN, and then embed it into the space of symmetric positive definite (SPD) matrices. Finally, matrix power normalization operation is employed to obtain more robust Gaussian descriptor. Experimental results on three texture databases demonstrate that RDGD is superior to its baseline B-CNN and the state-of-the-arts.