Nearest Neighbor Classification Improved Algorithm

As we know, classification (or decision) is one of the most fundamental and important issues in the field of pattern recognition. And it is also a very active research subject in computer vision field. How to find an effective classification algorithm to improve pattern recognition rate has become a key issue.KNN, as a classic classification algorithm, has been widely applied to digital recognition, face recognition and so on. However, KNN also has its deficiencies; the improvement for KNN has been a hot research topic. There are many KNN variants.In general, these methods can be divided into two categories:improving the computational efficiency or enhancing the classification performance. In this thesis, we focus on the improvement of the KNN algorithms, main work is as follows:1) We review the existing pattern classification algorithms and outline the improved versions of the nearest neighbors algorithms;2) To overcome the weakness of the classical KNN, we present a K-NN mean algorithm. The presented method has been tested on the NUST603HW database and the CENPARMI database. The experimental results show that the K-NN mean algorithm outperforms the classical KNN,and it also performer better than the state-of-the-art local mean method as applied to unbalanced data;3) We present a KNN linear regression algorithm, which can take fully use of the structural information of the K nearest neighbors. The presented method has been tested on the Yale B, FKP, AR and ORL databases. The experimental results show that the proposed KNN linear regression algorithm outperforms the classical KNN, the nearest subspace method and the linear regression method.