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

Multi-level Wavelet-CNN for Image Restoration