Bottom-up And Top-down Object Classification

With the advancement of technology, methods in the area of Object Recognition constantly increasing. More and more feature descriptors were proposed. These descriptors are robust to spin and has a scale invariance. At present, there are many object recognition method for a specific class, but do not apply to generic object recognition. In the reality, one descriptor are not valid for all situation. How to put an effective combination of many features is a very challenging but also very practical problems. In the other hand, a lot of extra bottom-up features may be extracted in one picture due to the lack of prior knowledge. The top-down template matching only applies to object localization rather than the object classification. This article propose a hybrid framework for object classification based on top-down and bottom-up information.This paper analyses the shortcoming of both top-down and bottom-up methods and describes the solution to hybrid model and take some experiments to prove the correctness of the idea. Major work include:1)This article proposed a method to change the object template to a vector which account for the template. The idea is to change each Gabor element response to a probability that the edge exist. then catenate them to a vector.2)This article proposed a method to extract the PHOW descriptor based on the Active Basis. The main idea is to extract the descriptor drop in the convex hull of the Active Basis. And then use the traditional BOW methods to train the model.3)This article proposed a framework that use both Top-down and Bottom-up information through multiple kernel learning methods. After extract all the Top-down feature based on the Active Basis and bottom-up features. Train all the features through the multiple kernel learning machine.