The Apply Research On Handwritten Numeral Recognition Technology Based On Multi-Models

Handwritten Numeral Recognition is a branch of OCR (Optical Character Recognition). It can be applied in a various fields, such as data statistics, financial statements, bank check, census information, mail sorting, which requires the accuracy of the recognition results. In addition, the freely style of written numeral and the context is unavailable, so it is not so easy to improve the recognition accuracy.Firstly, the element characteristics are extracted for the advantages of high speed and simple algorithm. It establishes a handwritten numeral recognition system based on element characteristics. Training the recognition system with test samples, result shows that element characteristics are not good for the defective samples which result in similar characters recognition errors.In order to make up for lack of the recognition system based on element characteristics, the main works is focused on the new characters feature extraction, named concave line feature. Because the concave line feature can describe the shape and position of contour lines, it is senitive to subtle structural differences in character. The experimentation proved that the proposed feature vectors are produced with the predominance of the element characteristics for classification to classify the similar characters. However, the concave line feature describes the local information, so it is unavailable to the noise interference. The image which is "broken strokes" will not be well recognized.In order to make up for lack of the recognition system based on concave line feature, the main works is focused on the new characters feature extraction, named rotation projection feature. It reflects the character stroke density distribution, solve the noise interfenrence problem. The experimentation proved that the proposed feature vectors are produced with the predominance of the concave line feature for classification and good tolerance to noise interference can be achieved for fusion recognition.Considering the complementary element characteristics, concave line feature and rotation projection feature, a multi-model is proposed. In the multi-classifiers fusion module, a new multi-models fusion approach which called weighted voting fusion algorithm based on prior knowledges was developed. The experimental results show that we can solve similar character recognition errors more effectively. The improving of the recognition rate of similar characters improves the recognition rate of one number, and improves the entire handwritten numeral recognition system.