Study In Fish-eye Image Correction Based On Straight Lines

Fish-eye lens has large field-of-view, which is one of the main reasons why it is critical in lots of applications, such as special photographing that have to get large sight region, and spheral-screen cine. However, fish-eye image usually don't obey our human's visual habit due to the nonlinear imaging model of the fish-eye lens. So it is always our wish to recovering it to be perspective image, which well obeys human visual habit. Methods for correcting fish-eye images are usually based on camera calibration theory, which need specially made high-accuracy calibration mould or molding board. Whereas, this thesis studied in the research subject of moving out fish-eye image distortion based on spatial straight lines. Those are the main works:Firstly, this thesis researched on the theory of straight-line-based fish-eye image correction. We all know, the deformation of general fish-eye image can be deemed to be nonlinear distortion on the general perspective image. The method on this thesis optimizes parameters of distortion removing model according to the basic principle that transforms curves, which should be lines but having been transformed to curves by fish-eye lens, to lines, using only scene information, and then get undistorted image via back-ward bilinear interpolation.Secondly, I implemented the basic algorithm of fish-eye image recovering with Matlab programming. And then analyzed the influences of varying quantity, length, distribution and coordinate span of the segments on the train set that are mapped from 3D line on the effect of distortion recovering after series of computer tests.Last but not the least, this thesis also researched on the distortion recovering problem of fish-eye images that have larger FOV, even be it pi. Then I found out that we generally can't recover a fish-eye image having a FOV of pi via the method on this thesis. The reason on theory is that the fish-eye image that have a FOV of pi gathers total visual information of a semi-sphere region, and when it be transformed to perspective projection image, the information on the edge of the efficient area of image would be mapped to infinite area. The reason on model is that it is hard for any of models to simulate the projection relationship of the points that should be recovered to finite points and that should be recovered to infinite points simultaneously, and the model would become instable if it is used to recover both sorts of points that would be recovered to finite points and that would be recovered to infinite points simultaneously.