Dense Contrastive Learning and Depth Dynamic Aggregation for Reference-based Super-Resolution

Reference-based super-resolution (RefSR) has attracted growing interest due to its capacity to leverage external prior information. RefSR involves migrating texture details from the reference (Ref) image to the low-resolution (LR) image based on their corresponding pixel or patch relations. Hence, high-quality matching and feature extraction are crucial. However, most recent RefSR approaches have primarily focused on advanced network architectures to enhance performance, often overlooking the potential of feature matching and extraction. In this paper, we introduce the dense contrastive learning and depth dynamic aggregation network for Reference-based Super-Resolution, namely DDSR, comprising a dense contrastive learning (DCL) network and a depth dynamic aggregation (DDA) module. To further enhance the accuracy of feature matching, we propose a novel DCL network that optimizes the feature space distribution in a dense contrastive manner, resulting in more precise feature matching. In addition, we design a DDA module that extracts features from the Ref image, utilizing variable features with multiple depths to extract feature information comprehensively. Experimental results demonstrate that DDSR outperforms state-of-the-art methods in both quantitative and qualitative measurements.