Effects of Degradations on Deep Neural Network Architectures

TL;DR

This study systematically evaluates six deep architectures under six common image degradation models using two diverse datasets, revealing that deeper networks are generally more sensitive to degradations while CapsuleNet offers notable robustness to noise. It introduces the NTT layer to boost resilience across architectures and proposes V-CapsNet to explore the benefits and limits of increasing depth within capsule-based methods. The results show significant gains in noise robustness from the NTT layer but reveal trade-offs with baseline accuracy and adversarial vulnerability, underscoring the challenge of achieving high accuracy without sacrificing stability. Overall, the work provides practical insights and architectural tools for deploying robust image classifiers in noisy or lossy environments, and it highlights directions for building inherently noise-robust networks.

Abstract

Deep convolutional neural networks (CNN) have massively influenced recent advances in large-scale image classification. More recently, a dynamic routing algorithm with capsules (groups of neurons) has shown state-of-the-art recognition performance. However, the behavior of such networks in the presence of a degrading signal (noise) is mostly unexplored. An analytical study on different network architectures toward noise robustness is essential for selecting the appropriate model in a specific application scenario. This paper presents an extensive performance analysis of six deep architectures for image classification on six most common image degradation models. In this study, we have compared VGG-16, VGG-19, ResNet-50, Inception-v3, MobileNet and CapsuleNet architectures on Gaussian white, Gaussian color, salt-and-pepper, Gaussian blur, motion blur and JPEG compression noise models.

Figures (12)

• Figure 1: Examples of image samples from the datasets proposed in this work. Top row: Synthetic digits dataset. Bottom row: Natural images dataset.
• Figure 2: Examples of different image degradations with gradually increasing noise. (a) Gaussian white noise, (b) Gaussian color noise, (c) salt-and-pepper noise, (d) Gaussian blur, (e) motion blur, (f) JPEG compression (quality).
• Figure 3: Comparison of classification accuracies of the CNN architectures against different image degradation models on the synthetic digits dataset.
• Figure 4: Comparison of classification accuracies of the CNN architectures against different image degradation models on the natural images dataset.
• Figure 5: Comparison of classification accuracies of the CapsuleNet architectures having different hyperparameter configurations against different image degradation models.