Research On Deep Network Handwriting Recognition For Smart Terminals

Smart phones and other intelligent terminals are more and more popular,the image processing and handwritten recognition with deep convolutional network methods achieved well on high-performance configuration server.However,due to the significant computational power consuming,it is hard to promote and deploy the methods directly as storage and computing resources limitation on smart terminals.To address the problems,essential methods of both rapid image processing and handwritten recognition with lightweight deep convolutional network is introduced in this paper.Firstly,to speed up the image processing on smart terminals,the improved image processing methods is adopted,such as rapid image compression method,the processing of large-scale color images can be accelerated.Grayscale image enhancement method,the contrast between target image and background image can be improved.Weighted average image denoising by Gaussian filter,the high-frequency noise can be removed effectively.Optimal image binarization by iteration,the target and background image can be segmented clearly.Image expansion corrosion improvement,the fracture in target image can be repaired obviously and isolated noise points is eliminated further.And image segmentation method,the real handwritten string image can be extracted quickly and effectively.With the improved method,the time consuming of image processing has been shorten to seconds significantly and efficiency of target object extraction has been improved.Then,to improve the recognition accuracy of the Le Net-5 model,the model improvement and optimization methods is adopted,such as to improve the ability of model recognition by adjusting network capacity.To improve the ability of local image feature extraction by designing an appropriate size and number of convolution kernel.To reduce model overfitting by adding the dropout layer.To find out the optimal solution of loss function by choosing the matched activation function and optimizer.And to improve the efficiency of model training by selecting the appropriate batch processing and learning rate.With the improved and optimized model,value of loss function has been reduced effectively,recognition accuracy of test handwritten digits has been improved to 99.36%.Finally,the practical application of handwritten recognition based on deep convolutional network for smart terminals,uses methods of continuous convolution layers combined with 3x3 small convolution kernel and dynamic learning rate.And adopt the optimization and lightweight methods of depth separable convolution.The perception and classification ability of local image feature changes is improved.More abundant image feature is extracted.The convergence speed of model training is accelerated.The storage occupied by the lightweight model for smart terminal is reduced to 0.9%,the total parameter is reduced to 0.7%,and the floating-point operations computation is reduced to 0.66%,which reduces the storage and computing resource of the model significantly.the time consuming of handwritten recognition is reduced to 56%,The test accuracy of the improved model for smart terminal is improved to99.63%.The edge padding ratio of handwritten digit image is set to 20%,which improves the recognition accuracy on the real verification data effectively.The model training test and verification separation methods is used to train and persist the lightweight model on the server.The model file is ported and applied on the smartphone.The mobile application program is designed,the real handwritten string image is preprocessed quickly,and then the target digit string recognition is carried out by inputting the model.Therefore,under the condition of limited hardware storage and computing resources of smart terminal,the research methods in this paper can still carry out the rapid processing of large-scale handwritten letter digit string image and lightweight porting of deep convolutional network model,which improves the recognition accuracy of target digit string effectively,and is conducive to the application and promotion on different intelligent terminal platform.