Image Classification with Hybrid Shearlet Scattering Networks

Image classification is a challenging problem since large variabilities like deformations and illumination changes are existed in the images. In this paper, we introduce a shearlet scattering transform which uses the cascade of shearlet transforms and modulus non-linearities to obtain invariant image representations. It improves the sparsity of the previous scattering transform using Morlet. Furthermore, we give a deep hybrid network for image classification. It uses the shearlet scattering transform as the initial layers of a convolutional neural network. Experiments demonstrate that the proposed method can perform better accurate classification results, especially in the limited samples setting.