Research On Sub - Pixel Edge Detection Algorithm Without Coarse Localization

With the development of science technology, people concern more in accuracy in production. Correspondingly the high-precision is always needed in detection field such as the detecting on small module gear and optical components. With the emergence of new technologies such as precision injection molding, the production is more and more efficient. But how to do sophisticated detection is still unsolved and the traditional detection methods had been difficult to meet modern detection requirements. The sub-pixel detection is needed.The traditional sub-pixel edge detection techniques need to do rough positioning before sub-pixel positioning. It is useless to do the precise positioning if the rough localization is not accurate enough. To solve this problem, this thesis proposed a non-rough positioning sub-pixel edge detection algorithm to get rid of the dependence on rough localization and do the sub-pixel edge detection directly, which can not only eliminate the time of rough localization but also greatly reduces the chance of making mistakes.Firstly this paper takes principal axis analysis method to do color image dimension reduction. By projecting color vector of each pixel onto the principal axis this method finishes dimension reduction, reduces the amount of computation and preserves most of edge information at the same time. Then this thesis proposes a median filter algorithm based on bisecting K-means. Bisecting K-means median filter algorithm needs to classify the pixels in a certain range into ordinary pixels, edge pixels and noise pixels to judge whether edge or noise exists before filtering. Then different filtering methods are taken according to the judgment. This algorithm can also enhance the image edges when filtering out the noise. Finally this thesis proposes the non-rough positioning sub-pixel edge detection algorithm. By searching for a series of points which comply with the fitting requirements and doing Gaussian fitting in the gradient image the sub-pixel location can be found. This algorithm gets rid of the dependence on rough positioning and strengthens the reliability of positioning result. Meanwhile, this thesis also verifies the independent of gradient direction in edge detection by simulation, which effectively reduces the complexity of gradient calculation.At last this thesis compares the algorithm proposed in this thesis and the traditional one by doing sub-pixel edge detection in an image. The sub-pixel coordinates gotten by this algorithm consistent with the traditional one, the detection accuracy is 0.01 pixel level i.e. 0.1um, and the operation time is about two-thirds of the traditional method. And because there is no need to do rough positioning the reliability of the algorithm is better than the traditional one. In summary, the algorithm proposed by this thesis is effective.