| In the21th century, people’s work and life have been changed greatly by the population of computer and internet. Computer comes to ordinary people; internet turns to be a necessity. As a core technique of Internet, multimedia techniques update the internet services again and again. Text based services no longer meet public’s needs.As a result, image based services pour into the world. It is applied successfully into multiple areas:from weather report to product defection detect, from community security to medical image analyze. Image tagging becomes a hot topic, which exacts information from the image itself. However, it is still in a beginning of this research, no matter template matching or Neutral Networks, or even supporting vector machine methods, all of which could only recognize few objects with low speed. The difficulty is with the increasing of objects, exponential computer cost is required. Besides, the interferences between classes also go worse. In order to solve these problems, this thesis implements and improves a method popular abroad, which named TextonBoost. This thesis proposes a new filter bank and a new searching strategy in finding best shared subsets. The edge and lightening information of component a and b in the original filter bank are cut off and combined into a new component H. Though gabor filter,4direction information are got instead of two.It is proved that the clustering time is shorten by14.5%under the same situation, besides, the distance between classes is shorten by almost50%.The new sharing set searching strategy introduces randomness, which increases the possibility of greedy search to become the global best solution,and makes almost all the solutions become good enough and time performance33%better. This algorithm improves the time performance and makes it possible to detect multiple classes in a given image in an acceptable time. |
