Streaming Networks: Increase Noise Robustness and Filter Diversity via Hard-wired and Input-induced Sparsity
This addresses the issue of CNN performance degradation under noise for applications like image recognition, representing an incremental improvement in network architecture.
The paper tackled the problem of robust recognition accuracy for noise-corrupted images in CNNs by introducing Streaming Networks, which use intensity slices and independent training to achieve higher noise robustness, with results showing that only the combination of hard-wired and input-induced sparsity enables this robustness.
The CNNs have achieved a state-of-the-art performance in many applications. Recent studies illustrate that CNN's recognition accuracy drops drastically if images are noise corrupted. We focus on the problem of robust recognition accuracy of noise-corrupted images. We introduce a novel network architecture called Streaming Networks. Each stream is taking a certain intensity slice of the original image as an input, and stream parameters are trained independently. We use network capacity, hard-wired and input-induced sparsity as the dimensions for experiments. The results indicate that only the presence of both hard-wired and input-induces sparsity enables robust noisy image recognition. Streaming Nets is the only architecture which has both types of sparsity and exhibits higher robustness to noise. Finally, to illustrate increase in filter diversity we illustrate that a distribution of filter weights of the first conv layer gradually approaches uniform distribution as the degree of hard-wired and domain-induced sparsity and capacities increases.