In this paper, we introduce a new perspective for improving image restoration by removing degradation in the textual representations of a given degraded image. Intuitively, restoration is much easier on text modality than image one. For example, it can be easily conducted by removing degradation-related words while keeping the content-aware words. Hence, we combine the advantages of images in detail description and ones of text in degradation removal to perform restoration. To address the cross-modal assistance, we propose to map the degraded images into textual representations for removing the degradations, and then convert the restored textual representations into a guidance image for assisting image restoration. In particular, We ingeniously embed an image-to-text mapper and text restoration module into CLIP-equipped text-to-image models to generate the guidance. Then, we adopt a simple coarse-to-fine approach to dynamically inject multi-scale information from guidance to image restoration networks. Extensive experiments are conducted on various image restoration tasks, including deblurring, dehazing, deraining, and denoising, and all-in-one image restoration. The results showcase that our method outperforms state-of-the-art ones across all these tasks. The codes and models are available at https://github.com/mrluin/TextualDegRemoval.

Semantic segmentation is a fundamental task extensively used in the analysis of high-resolution aerial images. Due to the objects with various scales widely appearing in high-resolution aerial images, the fixed square receptive field in most existing DCNNs cannot work well and usually leads to unexpected predictions. To alleviate this problem, we propose a light-weight scale-aware module (SWD), which is end-to-end differentiable and can be embedded in most existing networks. In our module, we employ the re-sampling scheme to make each element of convolutional patches adjust its position, which explicitly adjusts the receptive field size to fit different scales of target objects. Further, considering re-sampled feature maps as weighted maps implicitly introduces spatial attention mechanism. As a result, with only 0. 2M additional parameters, the network embedded with our module can automatically adjust its receptive field to better recognize objects in various scales. In the experiment, we evaluate our method on the ISPRS Vaihingen Dataset, especially the analysis of buildings category, which usually in different scales. We further compare it to the spatial attention mechanism and mainstream networks that utilize multi-scale information. The experimental results and comprehensive analysis demonstrate the effectiveness and efficiency of our proposed method.