Sub-pixel Accuracy for Extracting Corner Point Based on Support Vector Regression

As an important image feature, a corner takes significant position in camera calibration, pattern recognition and image matching area. A large amount of image corner points are the intersecting points of the edges of polygons. A corner point extracting method based on support vector for regression (SVR) was proposed aimed at extracting intersecting points. First, a digital image of geometric figures was collected with JAI CV-M4+CL array CCD device under natural lighting, and was transmitted into computer by image grabbing card X64-CLiProTM. Second, the original grey level image with noise was changed into edge information with single-pixel width after it was processed by noise reduction with edge-keeping filter, and then was edge detected with Canny operator and the contour was extracted. Third, regression function of each detected straight segment was obtained by training the SVR with the training point set, which had sub-pixel accuracy. The intersecting points of corresponding straight segments, which are exactly the under-detected corner points, were obtained by simple math works. Experimental results show that the proposed method for corner point extracting has a high accuracy and stability, and a strong robustness. Furthermore, all of the intersecting points can be extracted.