InMAS: Deep Learning for Designing Intelligent Making System

Abstract

Deep learning is one of the notable solutions when developing intelligent making systems (InMASs) for students' test papers and assignments to replace the workload of the teachers and educators. This paper recommends a design method of InMAS based on the You Only Look Once (YOLOv3) algorithm. Such a method can be used in carrying out experiments on algorithm problems and creating two dedicated datasets. The first is for localization and the second is for recognition. The YOLOv3 network is employed to identify the location and extraction of each mathematical problem in every image. In the recognition part, because of the low recognition rate of traditional optical character recognition (OCR) on the handwritten characters, the YOLOv3 network is widely used to identify the characters in each arithmetic problem. In the final step, the numerical operation is held on the output characters. The template matching approach is used to evaluate the arithmetic problems with the wrong operation in the original pictures. The experimental results show that the accuracy of localization is near to 1. The proposed method carries out favorably against Baidu OCR based on recognition accuracy, showing a high accuracy of 97.15%.