Research On Feature Detection Method Based On SUSAN Operator

Feature detection of digital images is a hot research issue in image processing. It is important for researches at home and abroad to research and discuss on feature detection. Edges and corners are the two most important features for human visual system, whose detected results will have a direct influence on the final image. Focusing on the corner detection and edge detection, the paper devotes into new methods on feature detection based on SUSAN operator, in order to gain a better accuracy, effectiveness and the robustness to resist noise pollution.The main innovation work of this dissertation can be summarized as the following:(1) While SUSAN corner operator can only detect corners in a single scale, a multi-scale SUSAN method on corner detection is presented which is based on Gaussian transform. This method employs the multi-scale property of Gaussian transform to create a Gaussian pyramid by implementing a different scales Gaussian transform to the original digital image. Then an improved SUSAN detector with an adaptive threshold is further employed to gain corner candidates in different multi-scales. Finally, after every candidate is relocated to a certain position in the original image, the real corners are selected with reference to their certain neighborhood information. Experimental results show that our method not only can detect corners effectively in different scales, but also is obviously superior to some existing methods in terms of the misdetection rate and accuracy rate.(2) Since SUSAN operator on edge detection requires a predefined threshold, an adaptive edge detection method based on SUSAN operator is suggested. The method calculates the USAN area via an adaptive threshold of template in the original image, Otsu method is used to compute the two thresholds of USAN field; finally the edge information is obtained by an edge response function. Additionally, a new objective indicator is proposed to evaluate the performance of detected edges. The experimental results show that the proposed methods not only can detect edges efficiently, but also have a certain ability to resist noise pollution.