Image Corruption Robustness

Image corruption robustness asks whether vision systems can keep recognizing, segmenting, or reasoning when blur, noise, weather, compression, lighting changes, or compound degradations alter their inputs. Common approaches use corruption augmentation, test-time adaptation, image restoration, or representations and losses designed to suppress nuisance variation; evaluations include suites such as CIFAR-C and ImageNet-C alongside medical, driving, and remote-sensing tasks. Work increasingly covers vision-language and foundation models, measuring severity, uncertainty, calibration, and downstream failures as well as accuracy. A key tension is whether gains on synthetic benchmarks transfer to real and unseen degradations, and how to improve reliability without sacrificing clean-data performance or masking confidence failures.