Liu Pengju ; Zhang Hongzhi ; Zhang Kai ; Lin Liang ; Zuo Wangmeng

Abstract

The tradeoff between receptive field size and efficiency is a crucial issuein low level vision. Plain convolutional networks (CNNs) generally enlarge thereceptive field at the expense of computational cost. Recently, dilatedfiltering has been adopted to address this issue. But it suffers from griddingeffect, and the resulting receptive field is only a sparse sampling of inputimage with checkerboard patterns. In this paper, we present a novel multi-levelwavelet CNN (MWCNN) model for better tradeoff between receptive field size andcomputational efficiency. With the modified U-Net architecture, wavelettransform is introduced to reduce the size of feature maps in the contractingsubnetwork. Furthermore, another convolutional layer is further used todecrease the channels of feature maps. In the expanding subnetwork, inversewavelet transform is then deployed to reconstruct the high resolution featuremaps. Our MWCNN can also be explained as the generalization of dilatedfiltering and subsampling, and can be applied to many image restoration tasks.The experimental results clearly show the effectiveness of MWCNN for imagedenoising, single image super-resolution, and JPEG image artifacts removal.