| The depth information plays an important role in the digital image processing (DIP) because we cannot reconstruct movements and features of the objects in the three-dimensional (3D) space if there is not any depth data. Therefore, in this thesis, three image processing algorithms are proposed based on the depth from defocus (DFD) to obtain depth information. DFD is a very important visual module in the field of computer vision. It was first introduced by Pentland in 1987. The DFD technique requires only two images acquired at two different lens settings to compute depth estimates.In general, implementing a DFD system involves two major tasks, namely, calibration of the DFD system, and development of an accurate depth estimation method. This work is mainly related to the second task. This dissertation achieves a numbers of outcomes as follows:Firstly, basic theories and methods of DFD based on image processing are studied in detail. In the algorithms DFD or image reconstruct are both based on the system's point spread function (PSF), i.e. the radius of the image point's blur circle.Secondly, based on Pentland's algorithm, transform image right angle coordinate series's Fourier transform to polar coordinates series's Fourier transform. The experiment results indicate that this method is feasible and the most sensing error is 4.3%.Thirdly, this dissertation introduces a dynamic referencing technique which does calibration of the value of w2+v2 by dummy blurring obtained by a Gaussian convolution. The improved method is independent of the texture of the object. In addition, the calculation of Fourier transforms at each image point is too expensive for a practical technique. Therefore, we need only the Fourier power according to Parseval's theorem.Finally, a new method for depth estimation from defocus based on moment-preserving is proposed. This method doesn't need point spread function of optical system. The experiment results show that this method is real-time and effective. |
PDF Full Text Request