A general model compression method for image restoration network

Abstract

Convolutional neural networks have achieved prominent performance in the image restoration field at the cost of massive network parameters and computations. Several model compression methods have been proposed, however, most of them are designed for high-level vision tasks, which originally have some tolerance for information loss. To make image restoration networks targeting for low-level vision tasks more compact and efficient while preserving comparable good performance, we propose a general model compression method. More specifically, a deformable convolution kernel and standard convolution factorization are proposed to compress the network. Then, symmetric dilated convolutions and attention mechanism are employed to compensate for the performance loss induced by former compression. The process can be regarded as a micro-to-macro network rebuilding. Extensive experiments conducted in three typical image restoration tasks demonstrate that the proposed method attains up to 8× network compression ratio while achieving comparable or even better performance compared to the original network.