Research On HDR Imaging Method With Dynamic Scene Based On Deep Learning

With the rapid development of science and technology and digital image technology,the demand for High Dynamic Range(HDR)imaging technology in daily life and scientific research has been increasing,among which the research of HDR imaging technology for dynamic scenes has received extensive attention.Traditional multiexposure HDR imaging methods can cause problems such as ghosting and information loss when dealing with dynamic scenes;the application of deep learning to HDR imaging effectively overcomes the shortcomings of traditional methods,but when facing poorly exposed dynamic scenes,serious ghosting problems still occur.In order to further solve the above problems,improve the de-ghosting and information recovery ability of HDR imaging algorithms in dynamic scenes,and generate higher quality HDR images,two different deep learning-based HDR imaging algorithms for dynamic scenes are proposed in this thesis.(1)A feedback mechanism-based HDR imaging algorithm for dynamic scenes is proposed.The method synthesizes HDR images based on a neural network with feedback structure,and gradually improves the quality of the synthesized images through iterative feedback.The feedback neural network is able to remember and store the information of the previous iteration,and the output of each iteration will participate in the calculation of the loss function,so the feedback information contains part of the information of the earlier reconstructed images,which enables the network to learn the features of HDR images earlier;the earlier reconstructed images tend to contain more ghost images,and this feedback information can help the network to remove these ghost images in the next iteration.The algorithm is capable of generating high-quality HDR images with more supervised signals and stronger early reconstruction capabilities than existing feedforward neural network-based algorithms,with improved performance in both ghost removal and information recovery compared to existing algorithms.(2)A dynamic scene HDR imaging algorithm based on LSGAN is proposed.The method proposes a densely connected module improved by a global branch as the basic component of the generator for feature fusion,which can acquire both local and global information of the images,so that the network can retain image details as well as better understand the overall features of the images and thus effectively identify and remove ghost artifacts.The method uses a two-stage training strategy,where the first stage uses content loss to pre-train the generative network with good HDR image generation capability,and the second stage adversely trains the generator and discriminator by LSGAN loss to further improve the quality of the generated HDR images.Compared with existing algorithms,this algorithm solves the problem of not fully utilizing the global information due to the limited receptive field,and has greater advantages in recovering image structure,removing ghost artifacts and improving the overall visual quality of images,and has a lighter network structure compared with the current GAN-based algorithms.In this thesis,the above two algorithms are compared with other classical algorithms in experiments on several publicly available datasets,and the results demonstrate that the two algorithms in this thesis have better performance in removing ghosting and recovering information to generate high-quality HDR images compared with most existing methods.