Based On The Dynamic Combination Of Multiple Classifiers For Handwritten Digit Recognition

The objective of multi-classifier combination is to make use of each classifier's good qualities in recognition performance and gains higher recognition rate than each classifier. Multi-classifier combination improves the recognition rate with complementarities among theses classifiers. In the traditionary combination methods the role of each classifier in combination is fixed. But in practical application different test samples' recognition reliability is different. One combination method can improve recognition rate for some samples but reduce it for others. To realize the best performance for almost all the samples we need to adopt different combination structure for distinct samples-dynamical multi-classifier combination.This text realizes Optimal Linear Combination method basing on recognition confidence. This method advances recognition confidence scaling recognition performance for each sample. In train phase divide train samples into different areas with their recognition confidence and apply OLC in different areas to get multi-classifier combination power vector. The samples in different areas use different power vector. This method can embody the capability difference that one classifier acts on different samples. In test phase we work out test sample's recognition confidence of each classifier and so gain the area and it's power vector it belongs to. Then we can combine multiple classifiers with this power vector. Apply this method to hand-written number recognition and the recognition right rate is improved notably than static OLC.After discuss Optimal Linear Combination method basing on recognition confidence, I research confidence theory farther and advance each-class confidence conception. A sample's each-class confidence reflects the reliability that one classifier recognizes it as each class. Basing on confidence's definition the confidence's measurement is uniform for different classifier. So each classifier's confidence is comparative. And we can directly use the each-class confidence as combination power vector. Compare with former method it's merits are simple, pellucid and this method can be appended new classifiers without affecting former training data. This method is applied in hand-written number recognition and experiment results also approve its feasibility.