Real time handwritten digital image recognition system based on edge computing equipment

Abstract

With the development of edge computing technology, real-time handwritten numeral recognition system deployed in edge computing devices has a bright future. However, the edge computing equipment has weak computing power and limited storage space, so the mainstream image recognition neural network can not guarantee the real-time performance on low-performance edge devices. To solve this problem, this paper designs a handwritten digit recognition system which is suitable for low performance edge computing devices. The system extracts the effective information such as the position of the number in the input image through the image preprocessing module, and infers through a lightweight neural network specially designed for edge devices. In this paper, the proposed handwritten numeral image recognition system is deployed on the edge computing device Jetson Nano. The experimental data show that the inference speed of our model is 10 times faster than that of the original Tensorflow inference, and 60 times higher than the neural network Mobilenet specially designed for mobile devices. At the same time, with the increase of input video resolution, the FPS of the system does not decline significantly, which can meet the needs of most edge tasks. Finally, the system also provides design and deployment experience for other edge AI tasks.