Off-line Handwriting Chinese Character Recognition Framework Design Based On Convolutional Neural Network

The handwritten Chinese character recognition has wide application prospect,such as mail sorting,office automation and intelligent input.However,because the Chinese character strokes are more complex compared with other common characters such as English letters and Arabic Numbers,and since the difference of personal writing style and habit,Chinese character becomes a great variety,it makes more difficult to recognize handwritten Chinese characters.Therefore,the handwritten Chinese character recognition has always been a research hotspot and difficulty.In this paper,the author aim to study the single character of the off-line handwritten Chinese character,and launch research on the existing problems of the off-line handwritten Chinese character,the specific domain knowledge and the characteristics of the Convolutional Neural Network?CNN?,the work is as follows:1.To solve the problem of stroke distortion caused by the difference of personal writing style and habit,we introduce a new network called Inverse Compositional Spatial Transformer Networks?ICSTNs?,and propose a new recognition framework for handwritten Chinese characters with ICSTNs and CNN.ICSTNs is based on the inverse compositional image alignment algorithm and can learn the parameters of homography transformation matrix according to the input images.Therefore,it can align and correct handwritten Chinese characters of various writing styles and stroke distortion.In the paper,we simulate our proposed framework on the CASIA-HWDB by TensorFlow deep learning framework.The simulation results verify that ICSTNs can align and correct handwritten Chinese characters,and improve the recognition performance of the whole network framework.2.CNN has significant effect on the recognition and classification of images,so in this paper,the author did a lot of experiments to research the structure and optimization of the convolutional neural network,including how the depth of the network,the setting of learning rate,the regularization technique?L₂ regularization and dropout technology?and Batch Normalization layer optimization algorithm.Although CNN has strong ability of feature extraction,it is like a black box and ignores some useful prior information in specific areas,which cannot be learned by itself.Therefore,the author used the traditional feature extraction algorithm such as Gabor and Sobel to extract the eight directional feature of the data samples and take it as prior knowledge.The author will fuse the feature map and the original data set as the input of whole network.Meanwhile,this method can expand the data quantity,which is good for solving the problem of lack of unconstrained data samples.Finally,the simulation results show that the traditional feature extraction algorithm combined with CNN can further improve classification accuracy of handwritten Chinese characters.