Reasearch On Low-Level Visual Feature Detection Based On Gradient And Phase Information

This thesis mainly focuses the research on the following three aspects: edge detection, straight line detection and video text detection, which are belong to the low-level visual feature detection. We proposed a novel feature detection model based on gradient and phase to implement edge feature detection reliably. To extract straight line and video text location information from the edge feature, we design and implement adaptive straight line detection algorithm and video text detection and localization algorithm respectively. The main contributions of this thesis are as follows:(1) An edge feature detection model based on gradient and phase (GP model) is proposed. This model achieves the combination of gradient and phase information. The use of GP model for detecting features has significant advantages over gradient-based feature detection methods which are sensitive to variations in image contrast. Meanwhile, our model can solve the "feature separation" problem which exists in phase congruency model. Moreover, our model can solve color image's phase congruency checking problem. Through the theoretic analysis and a large scale of experiments, we showe the GP model can implement low-level feature detection reliably, and can be well applied to visual information processing applications.(2) A novel Hough-based algorithm for straight line feature detection is proposed. According to the spatial resolution of the image and the direction of the line, the proper parameter resolution can be selected. First, the algorithm provides a coarse parameter resolution based on a low-resolution Hough transform which is used to initially identify approximate line direction. Second, flexible coarse-to-fine iterations will be performed until the accurate line direction parameter is obtained. It first analyzes the local parameter space at a coarse resolution and then zooms down into the vicinity of the peak at successive iterations, while the Hough transform is performed on a successively revised parameter resolution. The rapid convergence and accurate detection can be achieved benefit from the feedback strategy. Meanwhile we design a more comprehensive straight line end-points detection algorithm, which ensures to detect line segments in point set of varying discrete degrees at a high precision. (3) A novel gradient and phase based approach to video text detection and localization is presented. Through utilizing the adaptive threshold and the statistics coarseness feature of each horizontal (vertical) pixel line of difference image, the text location detection is carried out fast and effectively. To suppress the complex background interference, and perform the video text localization accurately and effectively, we detect the text key feature point based on the GP model in the local area. The proposed method is robust to various video image quality, background complexities and text appearances.These model and algorithms are applied to stamp digital museum project, and video text detection, and video quality analysis, which achieve satisfied effect.