Similarity Search and Extraction Network: A Novel End-to-End Trainable Reference Feature Extraction Module for Image Super-Resolution

This paper presents a novel end-to-end trainable reference feature extraction module termed the Similarity Search and Extraction Network (SSEN), which incorporates similarity-aware deformable convolutions. SSEN exhibits superior robustness and adaptability without PSNR degradation in the presence of irrelevant references. Moreover, the proposed method can be effectively utilized for exploiting self-similarity when no reference image is available, extending its applicability beyond traditional reference-based super-resolution (RefSR) scenarios.