In recent years, deep learning-based Video Super-Resolution (VSR) models have achieved remarkable success in reconstructing high-resolution videos. However, these models typically suffer from extremely high computational complexity and a massive number of parameters, which severely hinders their practical deployment on resource-constrained devices. To address this issue, this paper proposes a hierarchical pruning method for VSR models3 based on the L1 norm. By analyzing the structural characteristics of VSR models, the method designs a hierarchical differentiated pruning strategy, applying different pruning ratios to modules such as motion compensation, feature extraction, temporal fusion, and reconstruction. Specifically, conservative pruning is applied to critical motion alignment modules, while progressive pruning is implemented on the highly redundant reconstruction modules. Structural integrity is maintained in core components like the PCD alignment and TSA fusion modules. Experiments based on the EDVR 2model and validated on the REDS dataset show that the proposed algorithm achieves model parameter compression while keeping the loss in Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) within an acceptable range.
Liu Xuejian (Wed,) studied this question.