Investigating Noise Robustness of Convolutional Neural Networks for Image Classification Using Gabor Filters

Abstract

Alongside the ever-growing influence of convolutional neural networks(CNN) in the image classification task, its capability for working robustly with noisy images will become more critical. A practical classification system needs to deal with noisy images because the inputs from the real-world environment that such a system will have to operate in can be noisy. Inspired from findings in Neuroscience, this work investigates whether Gabor filters can benefit the noise robustness of CNNs. It is widely known that responses from V1 cells in the early visual cortex can be modeled with Gabor functions. Visual pipeline in mammalian brains has to learn how to extract salient features from a noisy environment throughout its lifetime. Thus, we can suspect that these Gabor filter-like responses from V1 cells contribute to the noise robustness of visual perception. Given the assumption that there are similarities between CNNs and human visual perception, we investigate whether using Gabor filters can enhance CNN's capability for classifying images infused with noise.