Recognition of handwritten characters on the sensor matrix using a neural network

Abstract

Nowadays, touch displays are becoming more common. This design combines an input device and an output device, and also allows you to create the most convenient interfaces for interacting with the user. Input recognition significantly expands the functionality of touch input, providing the ability to apply recognition of entered characters to launch a variety of functions and applications. Such systems existing today are not flexible enough in configuration and use. This paper presents the process of developing an infrared touch panel input data analysis system. A neural network is configured and trained to solve the problem of recognizing handwritten characters. The functional purpose of the program units has been briefly presented: graphical interface, COM-port reader, neural network. A graphical interface has been developed that visualizes the used touch matrix and displays function buttons, input and output fields. The interface consists of a main work window, a macro view and setup window, and an auto-identification timer setup window. A program block has been developed that allows you to create a neural network according to given settings, make a request to it and train it. Additional program functionality has been developed that allows you to run external files using the entered symbol, partially automating the program, helping the user. All developed modules have been functionally tested individually and the entire program together. Principles of neural networks training have been analyzed. Experiments have been carried out in which training of a neural network with different settings and training parameters is implemented. Among these parameters are: the number of epochs of learning, the speed of learning, the number of neurons of the hidden layer. As a result of the experiments performed, factors affecting the quality of neural network training have been identified. As a result of the training, a set of coefficients has been obtained that ensures stable operation of the neural network. The accuracy of the neural network is 96%, and the recognition speed is less than a second.