Research On 3D Shape Recovery Method Based On Single Image

As one important field of computer vision, 3D shape recovery shows more and more significant position and role not to be ignored in various directions of production and life in society and has wide application value in industry, agriculture, country defense, medicine, space technology, etc, especially. Shape from shading (SFS) is one of the critical techniques to shape recovery in computer vision, which makes use of the shading changes to recover such parameters as height value or surface normal to recover 3D shape.This paper adopts the common classification of SFS methods presented up to now, namely, minimization method, local method, etc, to each of which some typical algorithms are analyzed both from principles and experiments point of view. Comparisons and evaluations of these methods are also proposed in several aspects, such as the uniqueness of the recovered surface, the 3D shape recovery precision, the effectiveness and applicability of the algorithm, etc. Based on the analysis and comparison of various SFS algorithms, this paper develops a method of 3D shape recovery based on single image. The core idea of this research is using 3D cues left in 2D image-shading to recover 3D shape: (1) Deduce the conversion between object imaging coordinates and optical source coordinates; (2) Deduce the first differential coefficient of image gray level based on spherical surface hypothesis of local shape; (3) Advance a method of computing tilt under optical source coordinates; (4) Deduce the slant based on shading equation and spherical hypothesis; (5) Compute surface normal based on tilt and slant; (6) Obtain the surface normal under imaging coordinates based on the inverse conversion of coordinates ; (7) Compute the gray value and the height value.The "SFS 3D Shape Recovery Software"is designed in virtue of software development tool in MATLAB. Experimental results of a number of images prove that the designed method has relatively higher computing efficiency and recovery precision. The virtual reality sphere image is developed with "Third Dimension Image Developent Software". In the sphere image, actural gray value and real height value corresponding to each pixel are known in advance, which provide ideal initial condition for precision validation.The image of virtual reality sphere, whose size is 128×128 pixels, is selected as experimental object to analyze recovery precision in detail by applying the overall pixels error method, rectangle rim average error method and cross section average error methods respectively. Analytical results indicate that: (1) The maximum and minimum of overall pixels errors are 0.2586 unit and 0.1264 unit respectively; (2) The average error of pixels covered by the four rims of maximum inscribed rectangle reaches the maximun, namely, 0.2109 unit; the average error of pixels covered by the four rims of minimum central rectangule reaches the minimum, namely, 0.1264 unit. With the augment of sphere radius, the average error increases as well. (3) The maximum and minimum of cross section average errors are 0.2150 unit and 0.1674 unit respectively. Over 4/5 cross section average errors are below 0.2 unit.Because of sharp changing of gray level in the edge area, the errors are relatively higher. As for the area in which gray level changes gently, especially the central area of the virtual reality sphere, the errors are much lower and below 0.2 unit. With the accession of boundary condition and control condition of saltation part, and augment of image resolution, the 3D shape recovery precision will be further improved.