A multistage and multiresolution deep convolutional neural network for inverse halftoning

Inverse halftoning is a technology to restore a continuous-tone image from its halftone version. Recent works based on deep convolutional neural network (DCNN) have shown remarkable progresses in this area. However, it is still a hard work to accurately recover the content information, detail information and global information. To this end, we propose a multistage and multiresolution DCNN method for inverse halftoning. The network includes three sub-networks corresponding three stages, each of them is used to restore different information in a progressive manner. Firstly, a trTresolution analysis network (TRA) is proposed to remove halftone noise dots and then the initial reconstructed image is obtained. Secondly, the detail information is enriched by the detail enhancement sub-network through concatenating the initial reconstructed image and the input halftone image. Finally, a global enhancement sub-network is introduced to adjust information of the whole image. The evaluation results on three public datasets show that the proposed method is superior to the state-of-the-art methods in both visual quality and numerical evaluation. Moreover, the average runtime of the proposed network is 0.14 s for an image with the size of 256 x 256 pixels, which means the proposed network can meet the requirements of practical applications.