Because handwritten characters in practical bank check are always interfered by background image and seals,the accuracy rate of OCR is very low.To obtain required precision,most samples have to be rejected,by which the auto-processing rate is confined.In this paper,a new confidence evaluation method based on support vector machines is proposed.In this method,special SVMs are constructed for every character class to evaluate confidence and recognize similar character.This method makes full use of SVM's studying ability in the condition of few training samples,and can accurately estimate the confidence.By this method we can reject fewer samples to obtain required precision.