Research On Perceptual Grouping Of Image Edge

The main features of image include color, texture and shape."Shape"is very important in object recognition, which can be proved in human visual perception. In the mean time, the human visual perception ability can also provide reference models for the computer vision. By the way, the"shape"exactly means"edge"in practice. Nowadays, the Image Edge Detection technology has been developed, there are urgent requirements for introduction of new theories and methods. Thus, the combination of the Perceptual Organization technology and the Image Edge Detection technology will be a significant attempt undoubtedly.Perceptual Organization Theory is derived from the Gestalt Theory in psychology, and its mainly research objects are processes of perception and problem-solving. In the computer vision field, this theory utilize the human's perception laws, in other words, we can organize significant combinations or structures by the original data from vision system. Laws of the Perceptual Organization Theory mainly come from the Gestalt Theory. However, few of the laws can be applied in the computer vision field at present, reasons are complicated: first, it is too abstract to carry out for some laws; then second, Perceptual Organization Theory is a new developing technology.The Image Edge Detection method takes advantage of the gray scale alternation to identify edges. However, in spite of the noise sensitivity and the detection precision, these edges we got are only sets of the pixels, and can not be applied to the Perceptual Organization principle directly. Thus, in the research on perceptual grouping of image edge, we add a process of edge connection and straight line segments detection after the Edge Detection, and then organize and get the contour by the ways of the Perceptual Organization Principle. The main content of the paper is as follows.1. Detect edges of the original image: For the amount of the classical edge detection algorithms, we have a contrast test on the classical algorithms at first, including the Roberts, Sobel, Prewitt, LOG, and the Canny algorithm. Finally we choose the most effective method-Canny algorithm to extract edges, and get the edge-sets.2. Connect the edge-sets and change them into straight line segments: The Hough Transform is a low noise sensitivity and effective method for edge-connection and straight line detection. By analyzing the principle of the Hough transform, we propose an improved method, which makes a gradient threshold to select valid pixels and another gap threshold to segment straight lines. Thus, we can take advantage of the improved Hough Transform to receive candidate of straight line-sets. Experiment confirmed that the improved method reduces computational complexity and avoid the loss of small straight line segments.3. Perceptual Grouping: Through the two steps above, we only get the pixel-level edges, but it is far away from the object contour obtained by human vision system, not exactly the real contour. So we need to simulate human vision system to constitute some laws, such as Proximity, Similarity, Parallelism, Closure laws, and etc, then we find the final groups. In this paper, we choose some of the major rules and propose several probability models as the grouping laws. Finally, experiments confirmed its feasibility and effectiveness.