A unified neural MRA architecture combining wavelet CNN and wavelet pooling for texture classification

Abstract

This paper introduces a novel unified neural Multi-Resolution Analysis (MRA) architecture that uses Discrete Wavelet Transform (DWT) integrated Convolutional Neural Network (CNN) along with DWT pooling. As convolution with pooling operation in CNN has equivalence with filtering and downsampling operation in a DWT filter bank, both are unified to form an end-to-end deep learning wavelet CNN model. The DWT pooling mechanism is also used to further enhance the MRA capability of this wavelet CNN. Using the first two wavelets of the Daubechies family, we present here a comprehensive set of improved texture classification results with several updates in the model architecture. These updates in the CNN model architecture apply to any node generally associated with the time-frequency analysis of the input signal.