We propose TFkNN which is an introduction of test features to classifier combination of multiple kNNs with the aim of improving classification performance. In TFkNN, an unknown pattern is classified on each test feature by the k-nearest neighbor rule. Scores from each local classifier are then collected to decide final discrimination. We performed simulation to show that test features are effective subspaces. The proposed classifier was then demonstrated on artificial data having several overlapping conditions. The experimental results show that TFkNN can give better performance than kNN. TFkNN moreover can also stabilize the recognition rate from the different parameter k. Although we implemented TFkNN in series, it requires the recognition time not so more than the single kNN. We also applied TFkNN to a real character recognition, resulting a high performance.
展开▼
