Deformable CNN with Position Encoding for Arbitrary-Scale Super-Resolution

Implicit neural representation (INR) has been widely used to learn continuous representation of images, as it enables arbitrary-scale super-resolution (SR). However, most existing INR-based arbitrary-scale SR methods simply concatenate neighboring features and directly stack the position information with the image features, without fully exploiting the correlations among the input information. This processing method may produce artifacts and erroneous texture in the SR image. To address this problem, we propose a deformable CNN with position encoding (DCPE). Our method consists of three main components: (1) Deformable Feature Unfolding (DFU) module, which selectively concatenates the image features to ensure accurate recovery of texture; (2) Fusion With Learned Position Encoding (FPE) module, which generates position encoding that can be better fused with image features, thereby enhancing the correlation between them; and (3) Deep ResMLP module, which enhances the representation capability of the local implicit image function to focus more on learning the high-frequency information of the image, thus reducing the generation of artifacts in SR image. We conduct extensive experiments and demonstrate that our method outperforms previous methods in both qualitative and quantitative evaluations.