Parameterized Wavelets for Convolutional Neural Networks

Convolutional neural networks (CNNs) have become the prominent type of machine learning approach for visual pattern recognition but suffer from the tuning of a large number of parameters. In this paper, we propose and evaluate a novel parameterized wavelet convolutional neural network architecture for image classification with a far less number of trainable parameters. This model architecture mainly consists of two parts: parameterized wavelet decomposition layers followed by convolutional layers and fully connected layers. These wavelet decompositions comprise wavelet operations with learnable parameters that are updated during the training phase using the back-propagation algorithm. We evaluate the performance of the proposed architecture using several common image datasets and compare the results with influential shallow to deep CNN models. Our findings support the possibility of reducing the number of parameters in deep CNNs without significantly compromising its accuracy.