Research On Light Field Refocusing System Based On Guided Upsampling

Light field is an essential topic in the field of computational photography.The 4D light field was initially captured for realistic depth of field rendering,i.e.,refocusing.Image refocusing technique enables the photographers to change the depth of field after exposure so that they have more freedom in the selections of subjects to be highlighted.The refocused images derived from light field data usually costs much time and space,and with the development of light field acquisition equipment and super-resolution technology,refocusing algorithms are facing greater challenges.Therefore,this paper proposes to downsample the light field,calculate a low-resolution refocused image,and then upsample the refocused image to the original resolution to improve the time efficiency and reduce storage consumption.This paper focuses on the method to upsample low-resolution light field refocused images,and proposes two types of solutions.The main work of this paper is as follows:(1)According to the relationship between the angular difference and the defocus information after refocusing,this paper proposes a light field refocused image upsampling method based on weighted fusion.This algorithm evaluates the degree of blur for each pixel in the image by calculating the standard deviation of the macro-pixel image of the light field after refocusing,and then merges the high-resolution central sub-aperture image and the image interpolated to the same resolution to obtain the highresolution refocused result,according to the processed non-uniformity map.In order to evaluate the proposed method,experiments were performed on synthetic light field datasets and real light field datasets.The experimental results show that this method has lower costs in time and space requirements,and can derive a upsampled refocused image with sharp focused areas and blurred defocused areas.(2)To further improve the precision of the upsampled refocused image,a bilateral weighted guided upsampling algorithm is proposed based on the local feature of the difference between the light field refocused image and the central sub-aperture image.This algorithm models the guided upsampling task as a bilateral weighted ridge regression problem.The bilateral weight biases the model on the pixels that have similar intensity and position to the center of the local window.Thus the derived linear relation could better describe the relations between the patches of input and output difference images,the obtained result can be combined with the central image to achieve the purpose of upsampling of the refocused image.It can be concluded from the experimental results that the proposed method could derive refocused images with better image quality than existing methods when the sampling factor is increased.(3)Based on the above research content,a prototype system for light field image refocusing is designed and implemented,which can visualize different types of light field images and facilitate refocusing,the depth of field rendering effect can be changed by adjusting the focus position,and the efficiency can be improved by increasing the sampling factor.The proposed system is complete and stable,which verifies the practicability of the research in this paper.