Research And Application Of Segmentation Algorithm Of Billets In Complex Production Line Scene

Image segmentation is a key aspect in the image recognition process. The quality of segmentation result affects the results of follow-up steps directly, eg. Classification, tracking, image understanding and target recognition. At present, we have done some comparison and research on image segmentation with the purpose of extracting object billets from complex scene. Through comparison and analysis, we found that these conventional methods are very poor to extract objects from complex scene, especially the billets in complex produnction line scene. In order to resolve this problem, we employ the method of graph theory, and extract the objects through the way of human interaction.In this paper, we first introduce the basic concepts and algorithm of image segmentation, and then elaborate the basic theory of graph theory and the basic idea of graph cuts. Furthermore, we have summarized the merits and faults of some of the typical existing segmentation algorithms that based on graph theory. This paper will focus on the analysis and improvement of interactive segmentation algorithm based on graph theory, and proposes two improved algorithms with purpose of extracting billets from complex lighting scenes.The first one is the segmentation algorithm based on graph theory and optimal threshold model. Since present algorithms based on graph theory are time-consuming, and cannot extract billets from complex background perfectly. We combine the threshold method and the Graph Cuts algorithm; establish the optimal threshold model to compute the energy of segmentation. In this way, the algorithm will fast and effective. The second one is an improved segmentation algorithm based on Graph Cuts. In above threshold model, the samples that used to estimate the parameters come from the seeds that selected by users, so the quality of result will be affecting by the seeds. Moreover, in some gray images, some pixels in object have similar gray-level with that in the background, so it is better to adopt to color image segmentation algorithm. To solve these problems, we made the following improvements on the algorithm.(1) Selecting right color space through experiments.(2) Accelerating the speed and improving the accuracy of classification, through the introduction of improved K-means clustering algorithm.(3) Using the new method to compute the energy of segmentation, the property of color and distance can be express more accurately.(4) After initial segmentation, we correct the pixels that have been divided wrongly with connected region denoising and edge correction method instead of iterated method. The whole process will speed up.