Automatic Comparison Test And Identification Of Chinese Handwriting Based On Deep Learning

At present,convolutional neural networks have achieved good results in feature detection,but there are some shortcomings in the complex structural features of handwritten Chinese characters,such as the homomorphism between the internal structural features of Chinese characters.In response to such a deficiency,we use the method of capsule network to complete the feature extraction and classification based on the tracking data set.The capsule network represents the specific type of instantiation parameters by constructing the activity vector,and the activity vector is routed to the corresponding capsule in the next layer through the dynamic routing algorithm,so that the next layer of capsules gets a clearer input signal,making up for the deep convolution.The loss of information in the neural network pooling layer also captures more homomorphic features.The main purpose of this paper is to extract and classify Chinese character handwriting.The main work and achievements are as follows:(1)The identification of handwritten Chinese characters is not only the detection of local features of Chinese characters,such as the angle of inclination of the horizontal direction,the curvature of the vertical bending,etc.It is also important to have a better description of the relative spatial relationship between features.In this paper,a capsule network structure suitable for Chinese character handwriting identification is constructed.The capsule network uses vector to represent the instantiation parameters of Chinese characters,which can capture more Chinese character structure features.In order to improve the accuracy of the identification,the capsule network structure has been studied in depth:1)Using the traditional convolution operation data to build the basic capsule,use RELU to activate during the convolution process;2)abandon the operation of the Dropout layer to reduce feature loss;3)feature representation and transfer through the vector,in the original object entity characteristics Based on the feature attribute of the feature is added;4)The unique dynamic routing algorithm routes the feature data according to the similarity of the upper and lower capsules.(2)In the case of the same writing conditions,the effect of tracking acquisition is achieved by controlling the writing time.After preprocessing,correction and segmentation,a total of 104,800 handwritten Chinese characters are screened out.(3)For the Chinese handwritten texts with different degrees of adhesion,this paper proposes a method based on local projection to reduce the overlap of Chinese characters caused by global projection.(4)For the identification of Chinese handwriting,three different contrast experiments were used in this paper,and the comparison experiments with other four convolutional neural network methods on HWDB dataset and tracking acquisition dataset were given:1)Using multi-stream volume Accumulated neural network,convolutional neural network and improved capsule network were used to compare the accuracy of identification on HWDB dataset.The experimental results show that the improved capsule network identification accuracy rate is 95.82%,which is higher than the contrast convolutional neural network method;2)The algorithm generalization performance experiment was carried out on the tracking data set.The experimental results showed that the improved capsule network identification accuracy decreased by 1.43%,while the deep convolutional neural network method showed an average accuracy of 8.52%;3)Training samples In the sensitivity experiment,the method of gradually reducing the capacity of the training sample is realized.During the experiment,when the training sample capacity is reduced to 2.82×10~4,the identification accuracy of the other four methods shows a“waterfall”reduction.The accuracy of the identification of the network structure has been steadily decreasing.