基于改进Faster R-CNN的算式检测与定位  

Detection and location of formulas based on improved Faster R-CNN

作　　者：王巍 周庆华[1] WANG Wei;ZHOU Qinghua(School of Physics and Electronics,Changsha University of Science and Technology,Changsha 410004,China)

机构地区：[1]长沙理工大学物理与电子科学学院,长沙410004

出　　处：《智能计算机与应用》2022年第12期164-168,共5页Intelligent Computer and Applications

基　　金：国家自然科学基金(42074198)。

摘　　要：算术题批改是小学数学老师的一项重要任务。为了提高批改效率,可使用机器视觉的方法来检测和识别。算式检测与定位的准确性会影响后续的识别与批改结果,为了提高其准确性,提出了一种基于改进Faster R-CNN的基础算式检测与定位的方法。通过聚类分析数据集中算式的参数,对区域建议网络(Region Proposal Network, RPN)中锚框(Anchor boxes)的尺寸和比例进行了调整,减少了训练中的冗余计算,提高了收敛速度;同时用ROI Align替换ROI Pooling,避免了2次量化对检测精度带来的影响。实验表明,改进的Faster R-CNN提升了基础算式的检测定位效果。Correcting formula exercise is an important task for primary school teachers. In order to improve the efficiency of grading, machine vision methods can be used to detect and recognize. The accuracy of formulas detection and location will affect the results of recognition and correction. The paper proposes a method of basic formulas detection and location based on improved Faster R-CNN. Through clustering analysis parameters of the formulas in the dataset, the scales and ratios of the anchors are adjusted in Region Proposal Network, which would reduce the redundant calculation in the training to improve the speed. At the same time, ROI Align is used to replace ROI Pooling to avoid the impact of twice quantization. The experiments show that the improved Faster R-CNN improves the detection and location effect of basic formulas.

关 键 词：算式检测 更快的区域卷积神经网络 聚类 感兴趣区域对准 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]