Research On HDR Reconstruction Method Based On CycleGAN

With the rapid development of computers and digital images,High Dynamic Range(HDR)images show a broader application prospect compared with Low Dynamic Range(LDR)images,and HDR image reconstruction tasks have gradually become an underlying research hotspot in the field of computer vision.In recent years,neural networks have been applied to HDR image reconstruction tasks with their powerful feature extraction capabilities.There are two methods based on single-exposure LDR images and multi-exposure LDR images,the former generates HDR images with problems such as missing image details and inaccurate colors,the latter has alignment problems during the training process leading to artifacts in the generated images,and there is no large paired training dataset for LDR reconstruction tasks.In this paper,we propose a single-exposure HDR image reconstruction method based on Cycle GAN with the following three main research points.(1)To improve the image generation quality,the network structure of the Cycle-Consistent Generative Adversarial Networks(Cycle GAN)is enhanced.The Residual Network(Res Net)in the generator is replaced by the Dense Convolutional Network(Dense Net)to extract more input image features,and a multi-scale discriminator is used to input generated images of different sizes into the corresponding discriminator to comprehensively determine the image is true or false,prompting the generated images to be closer to the target domain images.(2)To solve the problem of missing texture details in the generated images,a Dual Attention Network(DANet)blending mechanism is added to the generator to form the DADB(DANet Dense Block)module.Different attention is given to different channels and pixel points in the generation network,and more attention is paid to the areas with less information such as overexposure and weak exposure,so that the image can also generate specific texture information in these areas.(3)To solve the problem of inaccurate color of generated images,the loss function is improved and Lcos＿lcs loss is proposed.lcos＿lcs loss includes cosine loss function and custom integrated loss function of luminance,saturation and contrast,which prevents the generated image color from shifting,improves the image contrast information and learns the color information of target domain images better.Finally,the improved model is applied to the task of HDR image reconstruction.In order to verify the effectiveness of the work done in this paper,the results obtained are compared with several classical algorithms.The experiments show that the model in this paper has excellent performance in various indicators,and has high robustness and generalization.It not only effectively solves the problem of missing details in the overexposed or weakly exposed area of the image in the traditional practice,avoids image artifacts,but also maintains the color consistency of the generated image and effectively improves the quality of the generated HDR image.