The Design And Implementation Of English Handwritten Input Method Based On Recurrent Neural Network

In recent years, with the continuous development of hardware and software technology, people have become increasingly dependent on mobile phones, tablet PCs and other intelligent terminal equipment. At the same time, as the touch screen of the hand-held intelligent terminal equipment is becoming more and more big, many users are not content to the input mode that can only identify a character at a time, but hope that the equipment can directly identify a word or even a sentence, so that can improve the efficiency of terminal equipment of handwriting recognition. Therefore, developing the handwriting input method that can run on the terminal device and identify the string is particularly important.The main work of this paper is to introduce the design and implementation of the English handwriting input method. Firstly, it starts with the basic development process of input method, and the user needs are analyzed in detail. Then according to the demand analysis, it designs and realizes the function of all modules. At last, the function test and performance test of the system are carried out, and the test results are given.The main contribution of this paper is to extend the function of the handwriting input method which is on the intelligent terminal equipment from character recognition to string recognition by the research and development of the English handwriting input method. English handwriting input method is divided into two parts on the implementation: model training system and client system. Among them, the model training system is based on recurrent neural network, at the same time, combined with the Long-Short Term Memory and connectionist temporal classification. The recurrent neural network is characterized by the existence of "self connection" between the same layer nodes. This special structure enables it to map the context information to the current output. Long-Short Term Memory is a special neural network node. The node contains a "cell", whose role is to remember the history of input information, so as to overcome the "gradient attenuation" problem. The dynamic programming algorithm is used for connectionist temporal classification, and the probability value of the output sequence of the target is obtained by calculating the forward and backward variables, which greatly reduces the time complexity of the algorithm. The main function of the model training system is to study out a good performance classification model according to a specific training algorithm on the basis of the training data set. This system is divided into three modules: The network initialization module is configured to neural network before training; the forward propagation module calculates the output probability matrix under the frame of RNN; the back propagation module corrects the weight value of the neural network according to the error term. The client system is based on the input method framework of Android system. Its main function is to use the classification model which has been obtained to identify the user input of handwritten information. This system is divided into two modules: Input method initialization module load identification model and generate the window; the text recognition module collects the user’s stroke information and process it.Through the actual test, it is proved that this input method can analyze the English string which is written on the screen of the phone by users, and can return the recognition result to the users. In addition, the performance indexes such as the time of training model, the recognition accuracy, the identify response time and the overhead of memory have all reached expected. All of them laid a good foundation for the on-line production of the input method.