Research On Key Technologies Of High Dynamic Range Imaging And Ghost Removal

High dynamic range(HDR)imaging can not only provide better visual expression,but also present the new and powerful driving force for the leapfrog development of digi-tal media.At present,image fusion based on low dynamic range(LDR)is still one of the main ways to obtain HDR imaging.However,during the process of image fusion,how to effectively remove the ghost effect caused by dynamic scenes,camera jitter,inaccurate image registration and large displacement of moving objects is the key problem,which is needed to be resolved in HDR imaging technology.The technical difficulties are summa-rized as follows: 1)The global based multi exposure fusion(MEF)algorithm is difficult to achieve the fusion goal of full light and dark details and rich colors close to the natural scene? The local multi exposure image fusion algorithm will increase the complexity of the algorithm.At the same time,the image details may be unnatural due to local over enhancement,or the halo and noise may be enhanced.2)Under the real dynamic scene environment,there are many complex and irregular motion forms,and the existing imag-ing technology means and methods are relatively single,which is difficult to effectively eliminate the influence of ghosts? 3)The existing deep learning methods provide some solutions for massive video image fusion,but it still needs to be strengthened in enriching the training data set and improving the generalization ability of deep learning algorithm.Focusing on the above key problems and technical difficulties,this thesis mainly studies the ghost removal strategy of the dynamic image moving region fusion based on large-scale moving scene and proposes HDR image fusion and ghost removal methods based on traditional and deep learning.The main contributions of this thesis are as fol-lows:(1)Aiming at the ghost problem in the process of HDR image synthesis under dy-namic scene,the source image sequence fusion with large motion displacement or large exposure difference is studied to remove the ghost influence,and a fusion scheme includ-ing image registration and image matching is proposed.A coarse registration strategy based on scale-invariant feature transform(SIFT)and a fine registration strategy based on normalized mutual information(NMI)are adopted for image registration.As the ro-bust image quality evaluation index,the structure similarity(SSIM)is used to replace the standard deviation or Euclidean distance for the objective optimization function of image matching.An HDR- SSP algorithm including image registration and image matching is proposed,which can retain the brightness,contrast,and other details of HDR images.It can effectively remove the ghost effect and improve the quality of image fusion.(2)A comprehensive implementation scheme of image registration strategy and match-ing strategy based on block structure is proposed.The registration algorithm based on mutual information(NMI)is embedded into the image matching algorithm based on struc-tural similarity(SSIM),and an integrated parallel HDR- RMP fusion model with image registration,image matching,and image reconstruction is constructed to ensure the im-age fusion quality and removes ghost effect.It effectively reduces the computational complexity.(3)Aiming at the problem of massive mobile video image fusion,based on the char-acteristics and advantages of generative adversarial networks(GAN),an HDR GAN image fusion algorithm is proposed.The loss function considering Wasserstein distance as the generator and discriminator is established to solve the problems of gradient disappearance and pattern collapse of the GAN network.The attention learning mechanism is introduced into the image synthesis module to effectively improve the detail and texture information of the generated image.The final synthesized HDR image has significant effects on image quality and ghost removal.Finally,the research work is summarized and the prospect of future work is pointed out.We will combine the algorithms of traditional and deep learning and exact the im-age pixels and spatial features into the optimization of the deep learning network,and a comprehensive image fusion scheme will be carried out.