Gradient-Guided DCNN for Inverse Halftoning and Image Expanding

Inverse halftoning and image expanding refer to the ill-posed problems which restore higher-bit images from lower bit ones. Many scholars have studied these problems so far, but the restored images still suffer either quantization artifacts or fine detail losses. Although recent deep convolutional neural network (DCNN) based methods have shown its advantage in these two problems, it is hard to restore high quality images with fine details if no extra information is feeded to the network. To solve this problem, this paper proposes a gradient-guided DCNN model for inverse halftoning and image expanding. The DCNN model consists of two stages. In the first stage, two subnetworks are designed to explicitly predict the gradient maps of the input image, which account for the detail information of image. In the second stage, the gradient maps, concatenated with the input image, are feeded to another sub-network to guide the reconstruction of the final results. Experimental results show that our method outperforms the state-of-arts in terms of both visual quality and numerical evaluation. In particular, our method better recovers the fine details of the images.