GuideIR: All-in-One Image Restoration with CLIP and Depth Guidance

Abstract

In low-level computer vision, image restoration is a fundamental task that focuses on reconstructing clean, high-quality images from their degradation. Traditional methods have used deep neural net- works targeting specific tasks such as denoising, deraining, and dehazing, achieving remarkable results. However, these methods require knowledge of the image degradation type for effective model selection during testing, which often leads to inefficiency and inflexibility. Hence, these methods fail to gen- eralize across diverse degradation scenarios. Recent advancements have introduced unified models to address all-in-one image restoration, though they face challenges related to disentangled representations and training efficiency. We introduce a novel unified framework for all-in-one image restoration em- ploying task-specific guidance and depth information. Our approach predefines degradation types and uses CLIP for zero-shot classification to guide the restoration network dynamically. In addition, by inte- grating depth information, our model achieves better restoration performance through object boundary identification and segmentation. Experimental results indicate that our approach achieves competitive performance overall and demonstrates significant improvements in deraining and dehazing compared to existing approaches.