Research On Middle And High Level Algorithm In Of Image Understanding And Its Application On RoboCup Middle-Size

Image understanding is a subject between computer vision and artificial intelligence, which researches on how to interpret image using computer system by imitating the biological systems of human beings. It gets information of objects in images mainly using image processing, and then combines with knowledge on specified scene to interpret the scene in the image. It is with important theories and wide applications. Since the Marr Vision Theory proposed by Marr in the 1980s, researchers get great progress in the area of image understanding.Image understanding has clear hiberachy, and the middle and high level algorithm involve region segmentation, object recognition and image interpret, which are very important. We do some researches on thechnologies in middle and high level of image understangding. In the part of image segmentation of middle level, although thousands of algorithms has been proposed, there is not a generic algorithm. The FCM algorithm(Fuzzy C-Means) has good performance when objects have fuzzy edges in image. In order to overcome its disavantage such as great loads, we proposed a modified FCM algorithm on HSL color space based on two dimension histogram. We used the two dimension histogram to get a good cluster number and cluster centers for FCM, then it can reduce the iterative times of FCM and make constringence fast. In the part of object recognition of high level of image understangding, traditional pattern recognition has to face two problems: the information from image is not sufficent and the object patterns are dubious.In view of the situation, we propoded attributed-graph for object recognition based on genetic algorithm, which combined the static/structre feature with prior knowledges on object positions in images, then constructed attributed-graph and fitness function, and serached the best result for object recognition through genetic algorithm.Robocup is robot world cup organized by the Robocup Federation,and the it has promoted the development of many subjects, such as mechanics, electrics, automation, pattern recognition, image processing, and artificial intelligence. The Robocup Middle-size is representative. The Middle-size soccer robot is composed of decision-making subsystem, vision subsystem, wireless communication subsystem and motion control subsystem commonly. As the only source of scene information, the vision subsystem is very critical, and it is a typical image understanding system. We described the image processing used in vision subsystem of robocup middle-size at aspect of three level processing thechnologies of image understanding: choose grey transformation and guass low pass filter for image pre-pressing; use the modified FCM we proposed above for image segmentation, and extracting objects' color and shape feature; adopting the simplified attributed-graph algrithm we proposed above for the object recognition, then transferring objects information to decision-making subsystem.The system had a good performance in the Robocup China Open 2007.