High Quality Image Deblurring Using Depth And Infrared Features

Image deblur is an important topic in digital image restoration.It can be divided into two main categories: non-blind deblur and blind deblur.Although non-blind deblur is simpler and more effective,it requires a known point spread function for the image,which is difficult to obtain for a real captured picture.Also,most of the non-blind deblurring algorithm can only deal with single kernel deconvolution problem,which does not fit the blurring distribution on real images.Blind deblur method mainly uses the single blur image only to estimate the point spread function and perform image restoration.Although these methods are suitable for real captured pictures,the lack of information for the latent clear image making all of the analysis based on the blurred image only.So the deblurring the results still need further improvement.Facing these problems,this paper proposed a new framwork for blind deblurring.By adding the depth map and infrared map as reference,The initial kernel is calculated more accurately,and can be estimated individually for different depth regions.Also,blur kernel is iteratively optimized according to the intermediate deblurring result,making it more close to the true motion trajectory.In the case of non-blind deblurring,a depth-based motion blur model is proposed in this paper.The model can realize multi-kernel blurring at pixel level,At the same time,it can achieve basic three-dimensional movement including translation,rotation and scaling.Aiming at the blur model,a corresponding deblurring algorithm is proposed.This paper also presents a depth restoration algorithm based on Markov random field to solve the problem of information lost in most depth images,The experimental results show that the proposed method can get the corresponding clear images for the objects with different depth values.In addition,the improvements in single-kernel deblurring method reducing the occurrence of ringing effects.Also,our non-blind deblurring algorithm,which restoring blur images at pixel level,is proved to be convergent and has resistance to noise.