Research On Inverse Halftoning And Regularization Algorithm Based On Deep Learning

Digital halftoning converts a continuous image into a binary image.Halftone technology is widely used in the printing industry.Inverse halftoning is the restoration of a continuous tone image from its corresponding halftone image,which not only remove halftone patterns,but also provide fine details and clear edges.It means a lot in practice.Inverse halftoning is an ill-posed problem.Inverse halftoning is a challenging but promising research topic.Deep learning is advancing very rapidly.More and more state-of-the-art algorithms about image processing are proposed.This paper proposes two inverse halftoning algorithms based on deep learning.GAN estimates generative models via an adversarial process.This paper proposes two inverse halftoning algorithms based on Generative adversarial networks in a supervised learning.1.We design a Generative adversarial networks which contains a dense block.It promotes the circulation of feature information in the network.It strengthens feature reuse and reduces the number of parameters.The algorithm is used to extract feature.It extract abundant feature information.2.We also design a Generative adversarial networks which contains residual blocks.The skip connection solves the problem of vanishing gradients and improves the performance of the feature extraction.It ultimately improves the image reconstruction accuracy.The algorithm is good at feature learning.Experiments are performed on error-diffusion halftone images and ordered dithering halftone images separately.Two inverse halftoning algorithms perform well.Two algorithms can improve the image quality and have a better reconstruction performance compared with other algorithms.Experimental results show that GAN does not have any problems with model collapse.The proposed method can effectively remove halftone patterns.The deep learning regularization algorithm has been studied.According to the characteristics of the existing deep learning regularization algorithms,the following improved algorithms are proposed.(1)Regularization approach based on Auto Augment and Spatial Dropout.(2)Regularization approach based on Auto Augment,Spatial Dropout and Drop Block.Finally,the improved regularization algorithms are applied to the baseline model which is Wide-Res Net-28-10.The regularization algorithm reduces the risk of overfitting.We evaluate our proposed algorithms on two object recognition benchmark tasks(CIFAR-10 and CIFAR-100).Valid information can be mined by analyzing and studying data.Simulation experiments are carried out by Tensorflow to prove the effectiveness of improved methods.Simulated results show that the neural network is effective in the prediction due to its high accuracy and good generalization ability.The experimental results show that algorithms proposed in the paper can obtain better results than the contrast methods.The improved regularization algorithms proposed in the paper can be widely used in other similar classification recognition problem.