Research Of Shape Detection Through Hough Transform Based On Local Analysis

Object recognition is an important field of computer vision, its production has an impact on industry automatization, space navigation, deep-sea exploration and national defence. Hough Transform is an important arithmetic in pattern recognition, an outstanding merit of Hough Transform is that it can transform the full-scale detection in image space to easy local peak value detection in parameter space. But the huge computing burden and storage space are the existing difficulties. In order to solve the difficulties of computing, storage and stability of Hough Transform to detect lines, circles, ellipses and some unregular shapes, the improved algotithms of Hough Transform based on local analysis are proposed.For the shape expressed by two parameters, for example, line and linar shape, a novel scheme that Hough Transform combines local principal component analysis(PCA) with parameter restrained is proposed. It makes use of direction analysis result of PCA to restrain the parameter space, it is valid in not only performance of speed and storage but the interaction difficulties which caused by different lines in image space. For the shape expressed by five parameters, for example, circle and ellipse, a method of hierarchical Hough ellipse detection with parameter restrained based on local PCA aiming at segments of interest and the result of curvefitting is proposed. This method which makes use of position information of edge image and directions of segments to exclude non-targets solved the problem of invalid sampling and accumulations. The 5D-parameter search difficulty is reduced, while the storage is controlled after the curvefitting.For the unanalytic shape, two fast algorithms of generalized Hough Transform to locate the target introduced by key feature points including directly voting and parameter restrained by the information of dominating points are proposed in this paper. The key feature points poss steady image information and the number is smaller than the image edge, which solve the difficulties of 4-D parameter space search burden and huge accumulator, and ensure the accuracy and speed at the same time.