Examining the Impact of Blur on Recognition by Convolutional Networks

Abstract

State-of-the-art algorithms for many semantic visual tasks are based on the use of convolutional neural networks. These networks are commonly trained, and evaluated, on large annotated datasets of artifact-free high-quality images. In this paper, we investigate the effect of one such artifact that is quite common in natural capture settings: optical blur. We show that standard network models, trained only on high-quality images, suffer a significant degradation in performance when applied to those degraded by blur due to defocus, or subject or camera motion. We investigate the extent to which this degradation is due to the mismatch between training and input image statistics. Specifically, we find that fine-tuning a pre-trained model with blurred images added to the training set allows it to regain much of the lost accuracy. We also show that there is a fair amount of generalization between different degrees and types of blur, which implies that a single network model can be used robustly for recognition when the nature of the blur in the input is unknown. We find that this robustness arises as a result of these models learning to generate blur invariant representations in their hidden layers. Our findings provide useful insights towards developing vision systems that can perform reliably on real world images affected by blur.

References in corpus (7)

• Very Deep Convolutional Networks for Large-Scale Image Recognition
• DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs
• Trunk-Branch Ensemble Convolutional Neural Networks for Video-based Face Recognition
• PsyPhy: A Psychophysics Driven Evaluation Framework for Visual Recognition
• Feedforward semantic segmentation with zoom-out features
• Faceless Person Recognition; Privacy Implications in Social Media
• On Classification of Distorted Images with Deep Convolutional Neural Networks

Cited by in corpus (34)

• AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty
• Generalisation in humans and deep neural networks
• Natural Adversarial Examples
• Benchmarking Robustness in Object Detection: Autonomous Driving when Winter is Coming
• Benchmarking Neural Network Robustness to Common Corruptions and Surface Variations
• Testing Robustness Against Unforeseen Adversaries
• Joint Face Image Restoration and Frontalization for Recognition
• A Study of Face Obfuscation in ImageNet
• It GAN DO Better: GAN-based Detection of Objects on Images with Varying Quality
• Learning Loss for Test-Time Augmentation
• Quality Resilient Deep Neural Networks
• A Study and Comparison of Human and Deep Learning Recognition Performance Under Visual Distortions
• Orthogonal Deep Neural Networks
• Improving Image Classification Robustness through Selective CNN-Filters Fine-Tuning
• When Human Pose Estimation Meets Robustness: Adversarial Algorithms and Benchmarks
• Contrastive Reasoning in Neural Networks
• Robust Semantic Segmentation with Superpixel-Mix
• Improving Robustness of Learning-based Autonomous Steering Using Adversarial Images
• A Self-Supervised Feature Map Augmentation (FMA) Loss and Combined Augmentations Finetuning to Efficiently Improve the Robustness of CNNs
• Deep joint rain and haze removal from single images
• Natural & Adversarial Bokeh Rendering via Circle-of-Confusion Predictive Network
• Towards Corruption-Agnostic Robust Domain Adaptation
• Defending Against Image Corruptions Through Adversarial Augmentations
• AQuA: Analytical Quality Assessment for Optimizing Video Analytics Systems
• Schrödinger's Camera: First Steps Towards a Quantum-Based Privacy Preserving Camera
• Noise Optimization for Artificial Neural Networks
• Untapped Potential of Data Augmentation: A Domain Generalization Viewpoint
• Robust Training Using Natural Transformation
• Frustratingly Easy Uncertainty Estimation for Distribution Shift
• Progressive Compressed Records: Taking a Byte out of Deep Learning Data
• Dirty Pixels: Towards End-to-End Image Processing and Perception
• Strategies for Robust Image Classification
• Linking average- and worst-case perturbation robustness via class selectivity and dimensionality
• Towards Robust Neural Networks with Lipschitz Continuity