Deep Learning Based Research Of Inverse Halftoning Model

In the field of image processing,inverse halftoning and image expanding are two classic problems,which aim to restore low bit depth images to high bit depth ones.Inverse halftoning and image expanding are inverse processes of halftoning and image companding,respectively.The technologies of halftoning and image companding represent grayscale of images through smaller bits,and are widely used in image compressing,storing and printing industry.However,this kind of image quantification process often result in irreversible image quality degradation,such as blocking artifacts,contouring artifacts and ring effects,which has a bad impact on visual effects.So far,there are many scholars have done research on the two problems of inverse halftoning and image expanding and have yielded some meaningful progress.Nonetheless,these methods are difficult to clean quantification artifacts and restore fine detailed textures at the same time.Based on deep learning technology,this paper proposed a gradient-guided progressive residual network for the weakness of existing methods.Our model can completely remove quantification artifacts while guaranteeing the quality of restored detailed textures.The main work of this paper can be divided into the following parts:1)To take full advantage of high level and low-level features in network,we choose U-Net as basic structure in the model,and propose a gradient-guided structure to utilize Sobel gradient information guiding network generate higher quality coarse restoration,which is helpful to improve the detail performance in a certain degree.2)Based on gradient guidance information,we incorporate the idea of residual learning in our model and propose residual learning structure.Fine tuning specific area of the coarse image generated by last step can be realized by learning a residual mapping,which further enhances the detail performance.3)Finally,the conclusion that image restoration results of the model outperform existing method both in numerical and visual aspects were demonstrated by a large number of experiments.Furthermore,ablative study is done to prove the reasonability and effectiveness of the model.