Research On High Dynamic Range Imaging Technology Based On Deep Learning

In today's digital age,high dynamic range(HDR)image can bring the public a more extreme visual experience.Current HDR imaging calculations can be classified into two categories:HDR imaging tasks based on single-frame high dynamic range(LDR)images and HDR imaging tasks based on multi-frame LDR images.After investigation,it is found that at present,for HDR imaging based on a single frame of LDR image,there is a lack of information in the brightly-exposed area,which leads to bright spots and artifacts in the reconstruction result;.The largest difficulty of the HDR imaging based on multi-frame LDR images is that the misregistration of multiple frames of images and the displacement of pixels cause artifacts in the final synthesized image.In response to the above problems,we tried to modify the existing HDR generation network,we add an exposure mask to the single-frame HDR generation network to deal with the bright exposure problem,and add a generative adversarial networks to the multi-frame HDR generation network to get more realistic results.Our work can be summarized as follows:(1)Aiming at the problem that the bright exposure details of a single-frame overexposed LDR image are difficult to restore and the quality of the result is low,we added the exposure mask branch on the basis of the expanded network,which is intended to enable the network to focus on learning the brightly exposed area.The result is more realistic and the possibility of bright spots is greatly reduced.(2)In order to generate a clearer HDR image,this thesis adds the perceptual loss function and texture loss function to the model training process,the purpose is to enhance the clarity of the reconstruction results and make the texture details of the image more real.(3)Aiming at the problem of artifacts in the synthesis process based on multi-frame LDR images,this thesis attempts to use the generative adversarial networks for multi-frame HDR imaging tasks,and after many ablation experiments,chooses the most suitable relativistic loss function to assist the learning of the generator.The experiments verifies that the generative adversarial networks is helpful to the HDR imaging performance,and our result is closer to the real image in color than others.In addition,it is unexpectedly discovered in experiments that using the channel attention module can further reduce the possibility of artifacts during multi-frame HDR imaging.(4)In order to give different degrees of attention to the different channels of the feature map in the generor,and to improve the clarity and color consistency of the image,we add a channel attention mechanism to the model.In the experiments,we also find that using the channel attention mechanism can further reduce the possibility of artifacts during multi-frame HDR imagingFinally,we conducted experiments on multiple newly added modules,and compared the two reconstruction networks proposed in this paper with others qualitatively and quantitatively.From the results,our improved results are more realistic than the baseline,and the visual effects are even better.