Investigation On Corner Detection Algorithm And Evaluation Technique Based On Image

Corner points are one important kind of image features in the field of image analysis and understanding. They are invariant under rotation, scale and translation transformations, and are stable under the change of light intensity. Corner detection can extract important information from images, with which the amount of data to be stored and processed can be significantly reduced, and the efficiency of subsequent image analysis algorithms would be improved. In this paper, after a comprehensive investigation of the existing corner detection algorithms, we propose a more effective corner detection algorithm and design a fuzzy evaluation technique for corner detection algorithms.In the first part of the paper, we investigate some existing problems of the multi-scale corner detection(MSCD) algorithms. This kind of algorithms calculates the curvature of contour shapes at several scales, and then recognize corners with a threshold filtering of the geometric mean of the curvature values. However, there are two potential problems of this kind of MSCD algorithms due to the use of the geometric mean of curvature values. The first problem is that the information of concavity and convexity of every corner is lost. The second problem is that the whole computation process is numerically instable. To overcome the difficulties, we propose a new kind of MSCD which is based on an arithmetic mean of the k-cosine digital curvature across multiple scales. Numerical experiments are carried out to evaluate the efficiency and stability of the new algorithm. Compared to the existing MSCD and other kinds of corner detection algorithms, the new algorithm has a good performance in terms of ‘Precision’ and an excellent performance in terms of ‘Recall’ and ‘Accuracy’.In the second part of the paper, we investigate the evaluation problem of corner detection algorithms. The existing evaluation technique is generally based on four basic data, including the number of ground-truth corners, the number of detected corners, the number of right-detected corners and the number of miss-detected corners. However, a serious problem is that the ground-truth corners are recognized manually with a binary decision at present. Researchers usually build up their personal ground-truth corner set. As a result, the evaluation results lack of objective criterion seriously. To solve this problem, a new evaluation technique of corner detection is proposed, which is based on the concept of fuzzy set. A corner membership value in [0, 1] is assigned to every image point, which can be decided according to a wide investigation of a crowd. The new technique can provide an evaluation result which allows the difference between individual visual concept and reflects the consciousness of human group. Compared to the existing binary evaluation technique, it is more consistent with the human beings’ visual system. Ten state-of-the-art corner detection algorithms are finally tested and compared with the new evaluation technique.