Research Of The Local Invariant Feature Detection Using Contour

Local invariant feature detection methods, which serves as the research focus of image processing, is the research fundament in many fields, for example image processing, image understanding, computer vision and pattern recognition. Since the images are always suffered from a series of transformations, such as rotation, scale, viewpoint, lightness, blur and so on, the issue that how to detect stable and unique local features is the research difficulty.Local invariant feature detection methods can be divided into methods using intensity and methods using contour by the processing information. Currently, there exists a lot of study on the methods using intensity, but less study on the latter. For the images with rich and stable contour, local invariant feature detection method using contour shows well performance. So it is very necessary to broaden the latter type of methods. In this paper, some theories about invariant features were deeply analyzed and some feature detection methods that using contour have been studied. With respect to the shortcomings of the existing methods, several novel methods with better performance have been proposed. The work in this thesis is as follows:①This paper deeply analysis the LoG(Laplacian of Gaussian) corner detection method. With respect to the shortcomings of unstable to noises for the LoG method, an improved method is proposed using the idea of multi-scale production which can effectively enhance response of corners and at the same time exclude noises. The experimental results show the better performance for the improved method.②With respect to the shortcomings of high time complexity for EBR(Edge Based Region) invariant feature region detection method and the poor affine invariance for the LoG invariant feature region detection method. A novel method for detecting local invariant features directly using contour is proposed based on the detected corners. This method mainly uses contour corner, angle bisector of corner and an invariant feature point relative to angle bisector. Because of the angle bisector is robust to noise and affected small by the rotation and scale, regions detected by the method has good stability, repeatability and low time complexity.③The repeatability of feature detection is the most important indicators to measure the performance of invariant feature detection methods. This paper also uses this indicator to test the proposed invariant feature region detection method using contour. The test images including kinds of transformation, such as rotation, scale, affine, brightness and noise. These changes are broadly representative in practice. The proposed method gets a high score of repeatability under these changes. The results illustrate that the proposed method has the properties such as fast, better robustness and wide range of applications.