Research On Image Highlight Removal Based On Deep Learning

In the field of image processing,the quality of images often determines processing results.However,in production and life,image shooting is affected by various factors such as hardware and environment,and the quality is often unsatisfactory.Among them,highlights will cover the texture,color and other information of the object,which is one of the main reasons for the dramatic decline in image quality.Therefore,this dissertation studies the causes of highlight and effective methods about highlight removal.In this dissertation,the basic theory of highlights is introduced firstly,and the research status and methods of highlight removal are summarized.Two highlight removal models based on deep learning are proposed.The main research contents and results of the thesis are as follows:1.The traditional methods predict the maximum diffuse chromaticity of an image to remove highlight components,based on the dichromatic reflection model.These methods require manual extraction of features,which is cumbersome and inefficient,and may result in color distortion.Since the popular deep learning has a good performance in the field of image processing and no highlight removal method based on deep learning has been developed.This dissertation proposes a divide-andconquer multi-resolution deep network to remove highlight.The method extracts the high frequency and low frequency information of the image patch by Laplacian pyramid to obtain the residual information of the highlight.Then,feature extraction is performed through the depth residual subnetwork to predict the maximum diffuse chromaticity.Finally,the removal of the highlight component is performed by the dichromatic reflection model.The results show that the method can achieve good results,and the image color is not distorted.Compared with other algorithms,our method is more effective for simple texture images.2.This dissertation proposes an end-to-end highlight removal model based on U-shaped structure dilated residual network.Firstly,U-shaped structure is used to extract texture and edge features of images with good edge recognition ability in a small-sized set.Then,the feature and the original highlight image are fused as the input of the dilated residual structure,and the highlight components are separated to obtain the diffuse component of the image.The results show that the proposed model can remove highlight effectively and has characteristics of strong color fidelity,less time consumption and strong real-time performance.Compared with the previous model and other algorithms,the average PSNR value of the method is improved.