Research On Methods Of Acquiring HDR Images Based On Deep Learning

High Dynamic Range(HDR)images have rich color and detail information,which can represent the actual scene more realistically.However,using hardware to directly capture HDR images is expensive and difficult to popularize.At this stage,most of them are obtained by software algorithms,including multi-exposure HDR image acquisition method and single-exposure HDR image acquisition method.The multi-exposure method needs to deal with the ghost effect caused by the foreground movement in the image and perform multi-frame alignment,while the single-exposure method can directly avoid such problems.This thesis mainly studies the single-exposure HDR image acquisition method based on deep learning.For the single-exposure HDR image acquisition method based on single-frame images,it is necessary to focus on the information reconstruction of poor image quality areas,such as highlight suppression in overexposed areas,noise removal in underexposed areas,etc.,and make the network focus on these poor exposures.In this paper,the author collects the research data of previous scholars and high-quality HDR images on the Internet,and combines the characteristics of the image exposure area to get much data for training and testing.This paper proposes two single-exposure HDR image acquisition methods based on different network architectures.One is a single-exposure HDR image acquisition method that uses a multi-branch automatic codec combined with a mask of the key area.In the network structure,this method uses a network structure with a multi-branch encoder to pay attention to the details and global information of the image,and combined with the badly exposed area mask to make the network more focused on the information enhancement of the badly exposed area.The other is a single-exposure HDR image acquisition method based on a dual-branch Unet-type network.This method uses two branches to deal with the information reconstruction of the overexposed and underexposed areas respectively,and uses instance normalization operations to perform the overexposure area of the image,the method uses bilateral filters to suppress artifact bands caused by noises in HDR images.In addition,the two methods proposed in this paper both use the color channel difference in the HSV(Hue,Saturation,Value)domain to guide the network to extract the image color distribution characteristics and reduce the color difference.Methods in this paper are compared with several existing representative HDR image acquisition methods to verify the effectiveness of the method proposed in this paper.The results show that the method proposed in this paper can restore the texture and color information of the overexposed area,while filtering out the noise in the underexposed area to reduce the generation of artifacts;in addition,the method proposed in this paper also has certain advantages in objective evaluation indicators.