Congratulations to @Li Jiaming and @Zhou Yixuan for achieving 6th place (6/199 = Top 3%) in the CVPR NTIRE2024 Image Super-Resolution competition! NTIRE is a renowned workshop held annually at CVPR, promoting a comprehensive understanding and discussion among researchers about crucial computer vision tasks such as image restoration, enhancement, and manipulation. Specifically, it focuses on how to restore degraded image content, fill in missing information, or achieve desired objectives (enhancing perceptual quality, content, or the performance of applications processing such images).

Currently, many are using the Swin Transformer as the base model to try to enlarge the model's receptive field or design complex modules to enhance the model's feature extraction capabilities. However, we observed a common phenomenon: the intensity distribution of feature maps tends to increase with network depth but is suppressed to a smaller range at the very end. This may implicitly limit the model's performance upper bound, as spatial information is lost due to encountering an information bottleneck, leading to the need for designing very complex networks or modules to push the upper bound of SR models. Therefore, we proposed the Dense-Residual-Connected Transformer (DRCT), aiming to stabilize the forward process through dense connections, allowing SR models to "avoid" information bottlenecks. Compared to current SOTA models, it saves 33% of the parameter count while achieving performance breakthroughs!