Research On The Corner Detection Algorithm Of Digital Images

In digital image processing, the extracted image feature information is one of thebasic work. Because the corner is of varying intensity in all directions and rotationconstant and does not change with the change of the light intensity, and corner also has acharacteristics of small amount of information and data. In many features of the image,the corner detection algorithm has broad application prospects. A good corner detectionalgorithm plays a crucial role in many computer vision tasks such as image matching,object recognition, motion analysis, and panorama stitching machine navigation.In this thesis, we firstly analyze the classic corner detection, point out the strengthsand weaknesses of the various algorithms. And we find that the curvature scale space(CSS) corner detection uses the Canny algorithm to extract the edge. We also analyze theCanny’s three criteria and Canny algorithm, and we know in the Canny algorithm, anGaussian kernel is used to smooth the image. Such a smoothing operation will have twodisadvantages: some weak edge will be smoothed out and the localization accuracy ismuch degraded, and that will cause some weak corner to be smoothed out and thelocalization accuracy of corners is much degraded in the CSS algorithm.Secondly, the finite differences representing the first-order derivatives in the2×2neighborhood are used to get the image gradient magnitude and direction in the Cannyalgorithm. This method is more accurate in edge location, but it is sensitive to noises,loses many details of the real edge and detects some false edge information. So we usethe four directions of a3×3neighborhood to calculate the gradient magnitude anddirection, which improves the accuracy of the edge positioning, and also inhibits somenoise.Finally, the improved Canny algorithm is applied to the CSS corner detection, andthe method of removing the false corners is also improved, so that at last we get a goodcorner detection algorithm that has better accuracy and precision. The experiments showthat the improved algorithm is effective.