A novel search coding method for generic object recognition based on shared features

摘要

In this paper, we consider the combined problem of distinguishing classes from the background and from each other, and propose an improved framework based on the previous state-of-the-art approaches. In the process of building ECOC (Error Correcting Output Coding) matrix (also called as sharing matrix), we adopt an encoding rule of one-versus-all, and maximize Hamming distance in categories as far as possible through heuristic search in sharing-code maps (I.e., layer joint boosting). Then the final classifier is responsible for detection, and ECOC matrix for recognition. In order to make full use of the output of the final classifier and its corresponding ECOC matrix, the following measures are worth considering: Firstly, a logistic function of the output mentioned above is used for a posterior probability of each codeword. Therefore the identified class label is the one corresponding to the codeword of Maximum a posteriori (MAP). Secondly, a similarity measurement utilizing the confusion matrix is advanced to focus on the similarities between classes. Thirdly, for the purpose of adaptive adjustment in Hamming distance, we change the subsequent search coding method according to the confusion matrix until the training errors are convergent. The experimental results illustrate the effectiveness of the proposed approach.