Scale-aware feature modeling and dynamic sampling enable arbitrary-scale image super-resolution.

Single-Image Super-Resolution (SISR) has witnessed a significant shift from traditional methods to deep learning, leading to flourishing developments recently. However, effectively handling arbitrary scales (e.g., integer, non-integer, or asymmetric) with a single model remains a challenging task. Existing SISR models fail to adequately perceive scale variations during the extraction of resolution features for arbitrary scales, while their limited local receptive fields hinder the modeling of global image structures. Furthermore, due to insufficient consideration of the impact of scale factors and local feature diversity on the upsampling stage, the generation of non-integer scale super-resolved images often suffers from incoherent details or abrupt transitions. To address these limitations, we proposed a cross-scale dynamic arbitrary-scale super-resolution network (CDASSR-Net) from the perspectives of cross-scale dynamic feature extraction and adaptive upsampling. Firstly, we propose a scale-aware feature adaptation (SAFA) module that adaptively adjusts filters according to the scale factor. Meanwhile, the cross-scale feature fusion via skip connections is proposed to better accommodating the demand for multi-scale feature representation and mitigating the loss of detail information common in deep networks performing such fusion. Then, we design a locally-adaptive scale-aware upsampler (LASU) module, which dynamically generates filters based on the input scale information, enabling upsampling to generalize to arbitrary resolutions. Extensive experiments conducted on various benchmark datasets demonstrate that integrating the proposed modules into fixed-scale SR networks allows them to achieve satisfactory performance on non-integer or asymmetric scales, while maintaining superior performance on integer scales. The code is available at https://github.com/Zheng4x/CDASSR.