High Dynamic Range Image Fusion Algorithm Based On Deep Learning

As image signal processing enters the digital age,people have higher requirements for the resolution,accuracy,authenticity and other performance of the acquired images.Standard dynamic range images gradually can not meet the public’s requirements for image display,so high dynamic range(HDR)imaging technology has been rapidly developed.At present,HDR imaging technology occupies an important position in image processing and computer vision industries,and has also become a hard indicator for the industry,such as mobile phones,cameras,automobiles and other industries to promote their own display products.However,traditional image sensors cannot directly acquire HDR images,and professional HDR imaging cameras are too expensive and inconvenient to carry.After acquiring a low dynamic range(LDR)image through the sensor,the method of reconstructing the LDR image into an HDR image by using an algorithm has become the mainstream.There are two main HDR image reconstruction technologies: single-frame LDR image reconstruction HDR image and multi-frame LDR image fusion HDR image with different exposure.Both approaches present several implementation difficulties.Based on the single-frame LDR image reconstruction method,when a limit scene occurs in the input image and the details are missing,the generated HDR image theoretically cannot carry the scene information.For the method based on multi-frame LDR image fusion,the first is how to effectively fuse LDR images with different overexposure or low-exposure scenes,and how to effectively fuse LDR images to generate HDR images with perfect scene and information.The second is how to effectively solve the ghosting effect caused by the dynamic scene in different frames of LDR images after synthesis.Finally,there is also the problem of multiple frames and single frames,the dataset required for training the model is very scarce and difficult to produce in large quantities,resulting in weak generalization ability and poor performance of the trained model.Compared with single-frame LDR reconstruction,multi-frame LDR composite HDR images have more sufficient scene input information,and this thesis focuses on multi-exposure HDR image fusion technology.In this thesis,a multi-exposure HDR imaging model based on deep learning feature distillation fusion is proposed,and a novel Feature Distillation Transformer Block(FDTB)is proposed to de-ghost HDR fusion images at the level of image semantic information.Aiming at the retention of extreme scene information in the input image,this thesis introduces hop connection to combine multi-scale features to produce HDR images with sufficient details.The method presented in this thesis is compared with several existing representative HDR image synthesis methods on several different test sets.The experimental results show that there are certain advantages in the objective evaluation index of image and subjective image perception.And it has good performance on multiple datasets,indicating that the proposed method has good generalization performance.Finally,the effectiveness of the modules and methods proposed in this thesis is proved by ablation experiments.