Based On Improved BP Neural Network Connected Handwritten Digit String Recognition

Connected handwritten digits recognition is a very difficult research branch inoptical character recognition research areas, while the value of it in real-life is alsovery important, since a wide range of usage in real-life, such as zip code, statisticalreports, bank notes, etc. The biggest difficulty of it is, digits may occur adhesions,broken pen, overlapping phenomena due to the connection. And because the digitsdon’t have logic between context, and often got involved into financial area, so fordigital recognition, requiring a very high accuracy, and there is no doubt that theimportance and rigor of this research.The current research of this field is focused on connected handwritten digitssegmentation and neural network optimization. First, for the digits segmentation,segmentation methods currently are focused on the use of structural analysis of thedigits, such as contour segmentation method, section segmentation method, etc.However, even though the complexity of these algorithms is high, the split accuracy islower and the speed is slower. Secondly, for the neural network optimization,optimization methods currently are more focused on structural, but there is no solutionfor optimization the recognition for connected handwritten digits.In order to solve these difficult problems in connected handwritten digitsrecognition, and to improve the recognition rate, the research work on connectedhandwritten digits recognition include the following aspects:Connected handwritten digits segmentation: segmentation is the foundation of theentire string of numbers neural network recognition process, effectively digitalsegmentation will have great impact on improving the quality of feature extractionand enhancing the recognition in neural network. Research uses the reservoir principle,identify all reservoirs in the connected digit string, which the reservoirs are created bythe digits adhesions, then through the analysis and calculations of the reservoirs,numeric string centroid, closed-loop and other structural features, propose thesegmentation path based on these reservoir characteristics. After comparison withother segmentation algorithms, this new method not only improves the segmentationspeed and also improves the accuracy of segment.Extract optimal feature set: To reduce the neural network input, while reducingredundant input and noise, need to extract features of handwritten digital sample asinput vector. Extract features by Fourier coefficients, stroke density features, contourfeatures, projection features, the center of gravity and the ce nter of gravity momentfeatures, coarse grid characteristics and features of the first black dot position; filtering features’ outstanding feature advantages and disadvantages by using a singlemethod, constitute the optimal feature set.BP neural network activation function optimization algorithm: A new algorithmto update BP neural network activation function parameters during training byintroducing an adaptive gain update algorithm, during learning may appear "flat area",conduct an effective solution. Based on the traditional BP algorithm, automaticallyadjust the parameters, will speed up the convergence of the network. Hile comparisonand analysis with other optimization algorithms, this method has not only improvesthe recognition rate, the recognition speed, but also decreases the consumption ofmemory, and achieved good results.