| Depth estimation is a key problem in machine vision,and it has a wide range of applications in many fields,such as face recognition,autonomous driving,3D reconstruction,etc.However,traditional cameras cannot record the angular information of light,which limits the further development and application of machine vision.As a computational imaging technology,light field imaging can record four-dimensional information including the position and direction of light simultaneously by improving the optical system.With the rich information,it can realize the functions of refocusing and multi-view extraction,which provide more clues to realize the depth estimation of the scene.Among them,the light field camera has the advantages of easy portability,convenient operation and stable imaging,so it has been widely used in the research of light field-based depth estimation.Based on the basic principle of light field imaging,this thesis discusses and investigates the focused light field imaging technology by designing and building a focused light field imaging system and post-processing of the obtained light field images.In addition,the depth estimation method based on focused light field images is also studied in terms of geometric parameter calibration and focus evaluation function of the focused light field camera.The main research contents of the thesis are as follows:(1)Based on the basic theory of light field,the imaging principle of light field camera based on microlens array is theoretically analyzed and the formula is derived,and the advantages and disadvantages of different structured light field cameras and their rendering methods are systematically analyzed,which lays a theoretical foundation for the design and algorithm research of subsequent light field imaging system.(2)The main parameters of the focused light field imaging system are analyzed and designed,and the system is simulated and modeled by Zemax optical software with the designed parameters.The main components of the system are selected according to the designed parameters,and the actual system is built and pre-processed for dark current correction and microlens center calibration.Conduct refocusing and multi-view extraction experiments to verify the feasibility of the system design and its 3D information acquisition capability.(3)The calibration method of geometric parameters within the focused light field imaging system is studied.By analyzing the imaging projection relationship of the system,a calibration model based on the mathematical relationship between the same image point and geometric parameters is established.For the needs of the calibration model,a three-stage calibration plate is designed to realize the calibration in a single shot,which improves the calibration efficiency.The error factors of the calibration model solution are analyzed,and a multi-directional solution method is proposed to reduce the calibration error.Finally,an experimental study was carried out by simulation and the actual system.The maximum calibration error for each parameter in the simulation experiment was 2.76%,and the average error was 1.5%.In the actual system calibration experiment,the subjective quality evaluation of the rendered image is taken as the basis,and the results show that the edge of the image rendered by calibration parameters is more continuous and there are fewer artifacts.The feasibility and robustness of the calibration method are verified.(4)The depth estimation method based on the focus stack based on the focus light field was studied.Aiming at the traditional focus evaluation function that cannot be applied to the refocused images of focused light field,through the analysis of the principles of refocusing and its departure and blurring,a focus evaluation function suitable for refocusing image of the focus of Light Field is proposed,based on this,the depth estimation of the focus stack of the focus of Light Field is implemented.Finally,an experimental study was carried out with the public data set and the actual system.In the public data set experiments,the detection rate of the focused evaluation function proposed in this thesis was 100% in the depth estimation experiments,while the detection rate of the traditional focused evaluation function was 11.1%.The maximum error of depth measurement in the actual system experiment is 4.85%,and the average error is 4.25%.The effectiveness of the algorithm was verified. |
