Video Super-Resolution Framework With Second-Order Grid Propagation and Cross-Frame Non-Local Attention Mechanism

Video super-resolution (VSR) technology has recently garnered substantial attention in the field of computer vision. VSR aims to upgrade low-resolution video inputs into high-resolution outputs, which is particularly valuable in applications such as video surveillance, ultra-high-definition content creation, and medical imaging. However, existing super-resolution techniques still face numerous challenges in terms of detail preservation and effective utilization of temporal adjacent frame information. Most modern VSR methods primarily depend on information from temporally adjacent frames to accomplish target frame super-resolution reconstruction, but they often suffer from limitations like losing detail during complex motions and the challenge of effectively extracting and using critical information due to significant differences between adjacent frames. To address these issues, this paper introduces two novel mechanisms designed to enhance VSR performance: the cross-frame non-local attention mechanism and the second-order grid propagation module. Unlike traditional frame alignment methods, the cross-frame non-local attention mechanism enables super-resolution processing without alignment, thus improving the system's robustness against extensive motion in videos and ensuring more accurate high-resolution video reconstruction. The second-order grid propagation module, on the other hand, extends the scope of feature information gathered during the super-resolution training process, overcoming the limitation of obtaining information solely from temporal neighbors and widely propagating and utilizing feature information to boost performance. Through rigorous experiments, our method demonstrates superiority over existing state-of-the-art VSR technologies across various datasets, showing significant improvements in detail preservation and overall video quality and validating the efficacy and advantages of both the cross-frame non-local attention mechanism and the second-order grid propagation module.