Research On HDR Imaging Method Based On Deep Learning

In practical application,due to the expensive cost of professional equipment,highquality High Dynamic Range Images(HDRI)are often difficult to obtain directly.Thus,researchers have started to use software to convert Low Dynamic Range Images(LDRI)into HDR images.There are two main methods for generating HDR images: generation method based on multi-frame LDR images and generation method based on single-frame LDR images.Because of moving objects in the scene or misalignment between images,the multi-frame-based method is prone to artifacts such as ghosting,sawtooth,and other.The single frame-based method can avoid these problems,so it has gradually attracted the attention of more and more researchers.The research content of this thesis is mainly the method of HDR image generating from single-frame LDR image.In the process of single-frame-based HDR image reconstruction,there are two main issues that need special attention: high-light suppression in over-exposure areas and noise elimination in under-exposure areas.We consider that these two tasks are two separable sub-tasks in the HDR image reconstruction process,and the solutions are fundamentally different.It is not suitable to use the same set of methods to solve them.Therefore,in terms of data,a large amount of HDR images were collected from existing research and Internet,and the HDRI-LDRI data set production method was designed carefully.We perform special degradation operations on different exposure levels in the image according to the distribution of pixel values for network learning of HDR image reconstruction of over-exposure and under-exposure areas of the image.In terms of network structure,we first propose a self-encoding and decoding structure network for reconstructing over-exposure area information of HDR images.The network combines the over-exposure area mask extraction method to let the network focus on the learning of the target area information,and uses the Instance Normalization(IN)module to make different types of images get more personalized processing of the over-exposed areas.Then,we analyzed the under-exposure area of a large amount of image data to identify the noise problem that mainly affects the HDR image reconstruction results.This thesis introduces a new network branch to reconstruct HDR images of under-exposure areas separately,and proposes to use a high-intensity bilateral filter to reduce the noise of the extracted under-exposure area mask.It can effectively avoid the problem of artifact bands in the final output HDR image.In the selection of the loss function,we proposed calculating the L2 distance between the output image and the reference image in the hue channel in the HSV(Hue,Saturation,Value)domain as a loss function to measure the color difference of the image to guide the network to learn the distribution of pixel values.In order to verify the effectiveness of the proposed method,an ablation experiment of different network module combinations and a comparative experiment of existing HDR image reconstruction methods are performed in this thesis.Experimental results show that our proposed method effectively utilizes limited information to restore the texture and color of over-exposure areas,and suppress the noise of the under-exposure areas to avoid the appearance of artifact bands.In addition,in terms of comparison of objective index scores,our method also shows obvious advantages.