摘要
针对教室视频图像中人物目标的识别问题,研究了深度学习R-FCN目标识别架构的算法优化及实验。通过分析R-FCN目标检测网络的结构和代价函数,采用自适应非极大值抑制修正预测框置信度,利用在线难例学习方法并优化候选框参数,得到优化后的R-FCN人物目标识别模型。经数据集训练及测试,实验结果表明,在测试数据集DL2021c下的人物目标单类别识别准确率为89.52%,该实验模型可为学校相关管理工作提供数据参考。
Aiming at the problem of person target recognition in classroom video images,the algorithm optimization and experiment of deep learning R-FCN target recognition architecture are studied.By analyzing the structure and cost function of R-FCN target detection network,using adaptive non maximum suppression to modify the confidence of prediction frame,using online hard example learning method and optimizing candidate frame parameters,the optimized R-FCN person target recognition model is obtained.After data set training and testing,the experimental results show that the accuracy of person target single category recognition under the test data set DL2021c is 89.52%.The experimental model can provide data reference for school related management.
出处
《科学技术创新》
2021年第30期88-90,共3页
Scientific and Technological Innovation
基金
广西高校中青年教师科研基础能力提升项目(2020KY41013)。