Diffusion-Based Adaptation for Classification of Unknown Degraded Images

Classification of unknown degraded images is essential in practical applications since image-degraded models are usually unknown. Diffusion-based models provide enhanced performance for image enhancement and image restoration from degraded images. In this study, we use the diffusion-based model for the adaptation instead of restoration. Restoration from the degraded image aims to restore the degrade-free clean image, while adaptation from the degraded image transforms the degraded image towards a clean image domain. However, the diffusion models struggle to perform image adaptation in case of specific degradations attributable to the unknown degradation models. To address the issue of imperfect adapted clean images from diffusion models for the classification of degraded images, we propose a novel Diffusion-based Adaptation for Unknown Degraded images (DiffAUD) method based on robust classifiers trained on a few known degradations. Our proposed method complements the diffusion models and consistently generalizes well on different types of degradations with varying severities. DiffAUD improves the performance from the baseline diffusion model and clean classifier on the Imagenet-C dataset by 5.5%, 5%, and 5% with ResNet-50, Swin Transformer (Tiny), and ConvNeXt-Tiny backbones, respectively. Moreover, we exhibit that training classifiers using known degradations provides significant performance gains for classifying degraded images.