Abstract: In order to meet the needs of paperless offices and greatly improve word efficiency, it is necessary to research and implement a handwritten digit recognition system. Handwritten digit recognition plays and important role in large-scale data statistics and the financial business, such as industry annual inspection, population census, tax statements and checks, etc. This project proposes a new type of handwritten digit recognition system based on convolutional neural network (CNN). In order to improve the recognition performance, the network was trained with a large number of standardized pictures to automatically learn the spatial characteristics of handwritten digits. For model training, according to the loss function, the convolutional neural continuously updates the network parameters with the data set in MNIST, which contains 60,000 examples. For model test, the system uses the camera to capture the pictures composed of the images generated by the test data set of MNIST and the samples written by different people, then continuously processes the captured graphics and refreshes the output every 0.5 seconds. With the trained deep learning model, we got a recognition accuracy of 99.3% in test process. Good performance in this experiment shows that our system can automatically recognize the handwritten digital content appearing in the target area and output the content label in real time.
Keywords: Digit Recognition, Deep Learning, Convolutional Neural Network (CNN), MNIST data set, Real-time recognition, Image processing, Machine learning
|
DOI:
10.17148/IJARCCE.2023.12551
