Texture Classification using a Hybrid Deep and Handcrafted Features

Abstract

In this paper, we have proposed a hybrid descriptor for the texture classification task. The feature variables are extracted from the approximation coefficients of the image, through a combination of deep neural network and handcrafted feature. The AlexNet along with completed joint scale local binary pattern (CJLBP) is used for illumination, scaling, and orientation invariance description. The wavelet decomposition layer provides robustness against additive white Gaussian noise. The feature dimensionality is reduced by using Principal Component Analysis. We have evaluated our proposed descriptor on the images of Outex texture databases. The experimental results presented in the paper in term of classification accuracy show that our proposed descriptor outperforms state-of-the-art feature extraction scheme.