Research Of Contour Detection Algorithms Based On Local Cues And Global Structures

Entering the Information Age, computers today have become essential tools to replace the human work in majority of research fields. The demands of improving computers’ flexibility have been requested along with the enhancement of functions and performance. For example, transformation between different types of information can be completed by human languages, vision and audio, and it also can be spread by various combinations. However, it’s important for operators to hold professional knowledge in order to operate well complex computer machines because computers need to obey a series of command orders and program through computer languages. To get rid of complicated and strict computer languages and instruction rules, computers must learn to adapt to human habits. That means the ability of seeing, listening and even talking should be necessary capacities of computers. Among these capacities, computer vision is such an ability to make visual sense work for computers.It’s been a long time since computer vision decides to build an artificial intelligence system, which can contribute to computer’s understanding of natural images. Computer algorithms need to detect and extract features from images before having the ability to recognize and distinguish objects or background in natural scenes. As a foundational problem in computer vision and digital image processing, how to increase the efficiency and accuracy of detecting edges and extracting features information has become the priority goal for edge detection and contour detection. Contrasted to previous research in edge detection, researchers discover that people want more about the relations between objects and backgrounds rather than only detecting and extracting all the edges from natural images in details. From this perspective, typical edge detection methods, such as Sobel detector, Canny detector and ACO(ant colony optimization), are no longer meet the requirements above. The direction of developing edge detection has been changed into contour detection and boundary detection since 2000. Also, contour detection methods begin to consider more local features(brightness, color and texture) than only one brightness derivative of images. Proposed Pb algorithm in 2005 is such a method by analyzing local cues and optimizing the combination of them. This method provides a better detection results than Sobel and Canny, and eventually opens the era to contour detection and boundary detection. CRF method proposed in 2006 first takes global structure feature into account to combine with boundary curves detected by Pb detector. It builds a CDT map and a global continuity model to connect discontinuous boundary curves, and increase the accuracy of results. Min-cover algorithm sees the detection problem as a minimum cover problem that every region in an image is covered by objects and backgrounds, where curves distinguishing regions are the boundaries.The main research contents are as followed:1. Do research in the mechanism and drawbacks of typical edge detection such as Sobel detector and Canny detector. And provide their comparison and analysis of results.2. Based on the research of typical detectors, this paper proposes an improved edge detection method, which improves original ACO(DATS) by using adaptive global threshold and adaptive distributed local threshold. By subjective comparison and quantitative analysis, proposed method indeed detects more continuous edges of interesting features than original one.3. Do research in the mathematical basis of Pb detector, CRF model and min-cover, and make comparison of results on BSDS using PR curves.4. Propose a contour detection method combining the multiscale local cues and improved comprehensive sampling matting. This matting method will take objects with several human labels as input and generate trimaps to detect the global contours of labeled objects. This paper proves that combination of objects’ global contours and multiscale local cues will give both of salient contours and local details. Besides, showing only contour information of certain objects in natural images without local cues is also available and change objects by amending labels. At last, the quantitative analysis and comparison of results from the proposed method and other six methods will be provided on BSDS dataset and in PR curves.