Adaptive Hybrid Composition Based Super-Resolution Network via Fine-Grained Channel Pruning

Abstract

In recent years, remarkable progress has been made in single image super-resolution due to the powerful representation capabilities of deep neural networks. However, the superior performance is at the expense of excessive computation costs, limiting the SR application in resource-constrained devices. To address this problem, we firstly propose a hybrid composition block (HCB), which contains asymmetric and shrinked spatial convolution in parallel. Secondly, we build our baseline model based on cascaded HCB with a progressive upsampling method. Besides, feature fusion method is developed which concatenates all of the previous feature maps of HCB. Thirdly, to solve the misalignment problem in pruning residual networks, we propose a fine-grained channel pruning that allows adaptive connections to fully skip the residual block, and any unimportant channel between convolutions can be pruned independently. Finally, we present an adaptive hybrid composition based super-resolution network (AHCSRN) by pruning the baseline model. Extensive experiments demonstrate that the proposed method can achieve better performance than state-of-the-art SR models with ultra-low parameters and Flops.