Study On The Algorithm Of Depth From Defocuing Based On Monocular Vision

Because its device is simple and suitable for robotic vision system, three-dimensional measurement based on image has been a hot topic in computer vision for a long time. Binocular vision technology and radiation stereo vision technology are the most representative and has been studied extensively. In recent years, the algorithm based on defocusing objects depth becomes important gradually. It belongs to monocular vision, and does not exist the problems of feature point matching and occlusion which yet not be resolved in stereo vision,so that it has better availability in some areas.Now, the algorithm of the depth from defocusing (Depth from defocus, referred to DFD) requires two or more images, and it identifies the depth of objects surface according to the difference of same object defocus imaging.The binding of secondary imaging increases the complexity of monocular distance operation, and it limits its real-time applications.In fact, the current DFD algorithm does not use images'overall gray information association, but applies the principle of radiate stereoscopic vision.When the gray level distribution of two-dimensional image reconstructs three-dimensional shape of the object, it also provides an restrictions of substitution the requirements of secondary imaging, so that it makes the theory based on a single image monocular objects surfaces ranging possible.This paper combining the gray level information of image, studies on the explorative subject which bases on a single image of the defocus ranging.The main work of this thesis is to study the DFD of single image, and consists of two main parts. The first part introduces clarity evaluation function during the seeking to the point spread function, and get the point spread function by the way of recovery by original images; the second part is to use two-dimensional gray scale image information to solve a single image DFD which is difficult to determine the relationship between specific focus image plane and image plane position by gray gradient method.The main contents of this paper is as follows:1. Some more clarity evaluation functions in the automatic focusing are compared. In light of the requirements of monocular defocus distance,this paper mainly analyze the performance of local region image clarity-evaluation-function, and selects the proper clarity-evaluation-function via experiments.2. To describe defocusing blurring by the gaussian function defocus point spread function model. The clarity-evaluation-function is the optimization criterion. This paper seeks to local defocus fuzzy parameters of the surface at different locations through image filtering method. 3. In order to determine the specific imaging plane position of objects, this paper uses two-dimensional gray scale image information instead of the classical secondary imaging constraints in DFD. It determined two adjacent properties relative position by solving the constraint equations gray. On this basis, it determines the specific location of the object image plane combining these two points in the image of the point spread function and the imaging lens model. This is the main innovation of this paper.4.In order to achieve a stable solution of gray constraint equations, this paper proposes a method that makes a continuous surface of the object plane approximation of the local area, the so-called pixels within the region are thought in the same plane and have the same normal, so that it can easily solve gray equations. This assumption is not just for the solutions of equations to add mathematics constraints. The physical significance is:(1) the principle of monocular defocus ranging is based on the fuzzy surface level of local area to achieve the object distance measurement, so strictly speaking, point by point measurement can not be achieved; (2) from the view of application point, in most cases, the task of robot vision is not to reconstruct the environmental objectives accurately, but some operation which does not require dense measurement, such as obstacle avoidance, navigation, recognition. Therefore, the proposed algorithm in this paper is to sacrifice precision in exchange for more simple practicality. This is the significance of the study.5. It shows feasibility of the algorithm by experiments, and provides a new idea based on ranging from single images for robot vision.