Study On Deep Learning Algorithm For High Dynamic Range Image Synthesis For Large Displacement

High dynamic range(HDR)images are widely used in various fields of production and life due to their rich visual information.The mainstream method of obtaining HDR images is to synthesize an HDR image from multiple low dynamic range(LDR)images that contain complementary information.Currently,deep learning methods have shown excellent performance in the problem of multi-exposure HDR image synthesis.However,deep learning-based multi-exposure HDR image synthesis methods still face the following challenges: 1)Existing deep learning-based multi-exposure HDR image synthesis methods cannot effectively solve the alignment problem of moving objects in different exposure images,which can introduce artifacts in the reconstructed HDR images;2)The scalability of existing deep learning-based multi-exposure HDR image synthesis methods is insufficient and can only handle a fixed number of images,limiting their use in various scenarios.To address these issues,the main contributions of this paper are as follows:To address the challenge of aligning moving objects in different exposure images,this paper proposes Cross-Attention HDR Fusion Network(CAHDRFNet),a HDR image reconstruction model based on the Cross-Attention Transformer.The CAHDRFNet model decomposes the multi-exposure HDR image synthesis problem into the alignment and fusion tasks of multi-exposure image features.In solving the alignment task of multi-exposure images,the CAHDRFNet model uses a CrossAttention mechanism to explicitly model the alignment task between different exposure images,and uses a Gating mechanism to suppress features that are incorrectly aligned from further participating in the calculation.In solving the fusion task of multi-exposure image features,the CAHDRFNet model uses a conditional network mechanism to help the model reconstruct the detailed features of different brightness regions in the HDR image.To address the scalability issue of existing deep learning-based multi-exposure HDR image synthesis methods,this paper redesigns the structure of the CAHDRFNet model that is related to the number of images,and obtains the CAHDRFNet-share model,a HDR image reconstruction model that is decoupled from the number of input images.In the CAHDRFNet-share model,this paper proposes an attention head score mechanism to highlight the more accurate attention head calculation results,thereby improving the performance of the model in the multi-exposure feature sharing processing module.To effectively fuse features from different exposure images,this paper uses prior knowledge of multi-exposure image fusion to propose a Gaussianweighted fusion method that fully utilizes the complementary information between different exposure images to improve the reconstruction quality of the HDR image.In the experimental part of the CAHDRFNet model,this paper demonstrates the superior HDR image reconstruction capability of the CAHDRFNet model and verifies the effectiveness of various designs in the CAHDRFNet model.In the experimental part of the CAHDRFNet-share model,this paper shows the scalability of the CAHDRFNet-share model and the advantages of the decoupled model,and in the ablation experiment of the CAHDRFNet-share model,verifies the effectiveness of the relevant designs decoupled from the number of images.