Handwritten Digit Recognition Based On Multiplayer Neural Network Classifier

Handwritten digit recognition has become a key research project in recent years, andit is widely used in the documents, notes, letters, etc. The core issue of handwritten digitrecognition is how to improve the recognition accuracy. However, due to different humanfactors, handwritten digits change with a lot of randomness and they are difficult to beidentified. This paper focuses on a sub-project of “research and software design ofquestionnaire automatic identification and statistical system”, the design object is toresearch and implement off-line handwritten digit recognition with high accuracy.This paper systematically discusses the main aspects of handwritten digit recognition,including image preprocessing, feature extraction and design of classifier. In order toachieve accurate character segmentation, this paper puts forward a fast connectedcomponent labeling algorithm, and on this basis, a combination of multi-methods isproposed for the segmentation of handwritten numeral string. Then in the normalizationprocess, a scale-down method is proposed to maintain good connectivity of the binary andavoid excessive features missing. In the stage of feature extraction, we get a more excellentcharacteristic combination strategy through experiments. We put forward a betterextraction method of concave feature and design a three-layer hybrid neural networkclassifier. Convex feature is adopted by rough classification, then the results input to thesecond and third layers. This classifier can improve the accuracy of handwritten digitrecognition effectively.Analysis and experimental results indicate that the connected component labelingalgorithm is simple and faster than the current conventional labeling algorithms, and thatthe reducing algorithm for binary images maintains good topology and consume less time,especially when it processes images with strong connectivity. This method can be used toeither digital images’ normalization or scaling applications such as topographic maps,region diagrams, etc. The proposed multilayer hybrid neural network classifier can effectively reduce the error rate of handwritten digit recognition, and can reach more than99percent accuracy rate for USPS samples.