Application Of Deep Learning In Image Highlight Removal

Image is an important source for people to get information,and the image quality determines what kind of information can be obtained.In the actual activity,the image will be disturbed by external factors,resulting in poor image quality,and lighting is a factor that exists in the environment that has a greater impact on imaging.Under illumination,smooth object surfaces are prone to specular reflections and highlights,can cover the original object texture,color and other information.And it leads to image quality degradation and many image processing algorithms to produce erroneous results.For this reason,this paper discusses and studies image highlight removal.This dissertation first makes an introduction on the basic principles of highlight formation and summation on the current status of research on image highlight removal,and describes the development and models of deep learning.Then two methods of deep learning in image highlight removal.The main research contents and results of the paper are as follows:1.A highlight removal model based on improved pix2 pix is proposed.To improve the network feature extraction,an improved ACBlock network structure is introduced in the generator to extract features using the residual network,and position-aware loss and mean-squared error loss are imported in the adversarial loss function to enhance the image highlight removal effect.The experimental results on SHIQ data set show that the method can remove highlights better.The average PSNR of the method is 20.139 d B and the average SSIM is 0.769.However,the predicted image PSNR and SSIM values are not high and distortion exists in the color of some images is impaired.2.To solve the problems of the previous method,a highlight removal model based on the improved U-net++ network is proposed.The method first uses the ability of the improved U-net++network that can fuse different levels of features and retain edge and detail information to obtain a feature map by preliminary feature extraction of the image,and then uses DDCM-Net to fuse local and global contextual information in the image to obtain the predicted highlight mask,highlight layer map and highlight-free map in turn.The experimental results on SHIQ data set show that the method has a good highlight removal effect and preserves the information of color and texture better.Compared with the previous model and other algorithms,the method acheives better highlight removal effect with an average PSNR of 33.68 d B and an average SSIM of 0.985.