Memory-Oriented Structural Pruning for Efficient Image Restoration

Deep learning (DL) based methods have significantly pushed forward the state-of-the-art for image restoration (IR) task. Nevertheless, DL-based IR models are highly computation- and memory-intensive. The surging demands processing higher-resolution images multi-task paralleling in practical mobile usage further add to their computation memory burdens. In this paper, we reveal overlooked redundancy of propose a Memory-Oriented Structural Pruning (MOSP) method. To properly compress long-range skip connections (a major source burden), introduce compactor module onto each connection decouple pruning main branch. MOSP progressively prunes original model layers compactors cut down peak while maintaining high quality. Experiments on real denoising, super-resolution low-light enhancement show that can yield with higher efficiency better preserving performance compared baseline methods.