Alternate Approaches to Scattering Networks in Image Classification

Abstract

Scattering networks are a special class of convolutional neural networks (CNNs) which implement a windowed scattering transform in their initial layers while learning the rest of the parameters. For classification tasks requiring little data, scattering networks beat cutting-edge deep neural networks. When given a huge dataset, their performance is comparable to end-to-end trained networks, but they're better suited for real-time applications due to their lower latency. The use of a windowed scattering transform for tasks involving image classification on the CIFAR-10 dataset is examined in this paper. We replace the 2-D Gabor filterbank in the state-of-the-art scattering network with alternate filterbanks that provide better directional separation in the frequency domain. We also develop a trainable directional filterbank that uses data-driven directional filters in its construction. The directional filters are built using the weights of a 2D convolutional operator. We demonstrate the performance of the alternate approaches in supervised classification tasks and observe that the trainable implementation outperforms the traditional scattering networks. © 2023 IEEE.