基于CenterNet的小学生英文手写体区域检测

为了探索智能批阅小学生作业的可行性，以小学生英文手写体为研究对象，建立了基于关键点的CenterNet模型。首先，针对低显存环境下CenterNet模型的构造与学习，提出了一种新的以组规范化(GN)替换批量规范化(BN)的池化模块结构改造方案，得到了改造版CenterNet模型;之后，将改造版CenterNet模型用于小学生英文手写体区域检测，实现了基于深度学习的英文手写体区域检测。将改造版CenterNet模型与原始CenterNet模型和CornerNet-Lite基准模型进行检测比较。实验表明：2种版本CenterNet模型的英文手写体区域检测精度和平均召回率均高于基准模型的相应值，改造版CenterNet模型的AP_0.5值甚至可达到73.1%，比基准模型高出近6%；此外，相比于基准模型，改造版的CenterNet模型的漏检情况更少，并在一定程度上有效抑制了误检。改造版的CenterNet模型不仅检测性能优于原始CenterNet模型，而且其学习过程更稳定、收敛更快,这为小学生作业智能批阅方案的设计提供了有价值的解决途径。

Detection of English handwriting area for primary school students based on CenterNet

To explore the feasibility of intelligent workbook review for primary school students, a CenterNet model based on the keypoints was established with primary English handwriting as the research object. Firstly, aiming at the construction and learning of CenterNet model in the case of low GPU (graphics processing unit) memory, a new scheme for pooling module structure modification was proposed by replacing BN (batch normalization) with GN (group normalization), and a modified CenterNet model was obtained. Then, the modified CenterNet model was used for the detection of English handwriting areas of primary school students, and the application of English handwriting area detection based on deep learning was realized. The comparison experiments with the original CenterNet model and the CornerNet-Lite baseline model show that the accuracy and average recall rate of the two versions of CenterNet model are higher than those of the baseline model, and the AP0.5 value of the modified CenterNet model can reach 73.1%, which is nearly 6% higher than that of CornerNet-Lite model. In addition, compared with the baseline model, the modified CenterNet model can get less missed detection and effectively suppress false detection to a certain extent. The improved CenterNet model not only has better detection performance than the original CenterNet model, but also has more stable learning process and faster convergence. This provides a valuable solution for the design of homework intelligent review scheme for primary school students.