Approach For Off-line Handwritten Chinese Characters Recognition Based On Knowledge Transfer With Feedback

Owing to the diversity of writing styles,the different samples of a Chinese character,has different positions among different components.And the irregular adhesion between parts of Chinese characters make it difficult to improve the recognition rate of offline handwritten Chinese characters.In order to solve this problem,the traditional pattern recognition method proposes to normalize the Chinese character samples.However,it is difficult to solve this problem fundamentally because the overall topological structure of Chinese characters can't be changed.In recent years,the rise of convolutional neural networks has brought new vitality to the problem of off-line handwritten Chinese character recognition.When convolutional neural network is used to overcome the above problems,it is necessary to ensure that the network level is deep enough.However,when there are too many hidden layers in convolutional neural networks,the gradient dispersion and over-fitting phenomena will easily occur in the training network,which will lead to the decline of network performance.In order to solve the above problems,the main research work of this paper is as follows:(1)An idea of decollate the offline handwritten Chinese character samples is proposed to solve the problem that the recognition rate of offline handwritten Chinese characters decreases because of the position difference between Chinese characters components.And transformed the cohesion between the original Chinese character components into the edge noise of the sample,which is convenient for the processing of the sample noise.(2)Using the knowledge transfer to trains segmented offline handwritten Chinese character samples.Decomposition of a network into several sub-networks and one primary network.Using sub-network to train segmented offline handwritten Chinese character components,then integrating the feature information of each Chinese character component in the sub-network into the main network using knowledge transfer.(3)The theory of knowledge transfer to overcome the phenomenon of gradient dispersion and over-fitting is put forward and explained.Decomposing the network model,the network hierarchy is reduced,and the restriction of gradient dispersion on the training of network parameters is reduced.Training the main network with soft targets,increase the training sample size indirectly,overcome the over-fitting.(4)Constructing a Network Model Based on Feedback Knowledge Transfer.Using feedback,adjust the weight coefficients of each sample adaptively to a reasonable interval to improve the network training effect.