摘要
智慧考场是智慧校园的重要组成部分,准确、快速地检测考场中的学生状态,是智慧考场应用的基本任务和关键环节.标准化考场中的考生分布相对密集且成像尺寸差异较大,而现有目标检测算法未充分考虑真实考场的环境特征,很难精确、实时地检测出考生目标,加之大部分目标检测算法需对不同目标手工设计先验锚框,模型部署范围受限.针对以上问题,提出一种高效的无锚框全卷积目标检测模型.该模型采用全卷积网络对输入图像进行逐像素预测,在可能存在目标的区域回归其包围框.在该模型中,设计了基于级联注意力的特征增强模块,通过逐级细化修正特征增强特征图的判别性,有效地提高考生目标识别精度;另一方面,针对真实考场中大量交叠目标检测问题,提出了点监督机制,以进一步提升交叠多目标的识别效果;最后,在构建的标准化考场检测专用数据集上,对所提模型进行验证.实验结果表明,与当前最先进的目标检测模型相比,针对真实复杂的考场环境特征提出的基于级联注意力和点监督机制的全卷积目标检测模型的m AP指标为92.9%,检测速度为22.1 f/s,泛化能力突出,综合效果最优.
Smart examination classroom is an important part of smart campus,and accurately and quickly detecting students in the examination classroom is a basic task of building a smart classroom.However,due to the dense distribution and imaging difference of the examinees in an examination classroom,most of the existing object detection methods can not precisely detect all the examinees in real-time.Moreover,most of the object detection methods rely on predefined anchor boxes,which are lack of portability.Aiming at the above problems,this study proposes an efficient one-stage object detection model based on fully convolutional network,which is anchor-free,with a prediction on the input image in pixel-level.In this model,a feature enhancement module is firstly designed based on cascade attention,which can effectively enhance the discriminability of the feature map by gradually refining and modifying the features.Secondly,in order to enable the network to distinguish overlapping objects in the examination classroom,a point supervision mechanism is proposed.Finally,this study verifies the above model on the special dataset of standardized examination classroom.With the cascade attention module and point supervision mechanism,the proposed model achieves 92.9%in mAP at the speed of 22.1 f/s,and is superior to most the state-of-the-art detection models.Especially,for object detection in new classroom environments,the proposed model achieves the best results.
作者
田卓钰
马苗
杨楷芳
TIAN Zhuo-Yu;MA Miao;YANG Kai-Fang(School of Computer Science,Shaanxi Normal University,Xi’an 710119,China;Key Laboratory of Modern Teaching Technology of Ministry of Education(Shaanxi Normal University),Xi’an 710062,China;National Engineering Laboratory for Integrated Aero-Space-Ground-Ocean Big Data Application Technology,Xi’an 710129,China)
出处
《软件学报》
EI
CSCD
北大核心
2022年第7期2633-2645,共13页
Journal of Software
基金
国家自然科学基金(61877038,61801282,U2001205)
陕西师范大学研究生创新团队项目课题(TD2020044Y)
空天地海一体化大数据应用技术国家工程实验室开放课题(20200201)
关键词
目标检测
智慧考场
无锚框方法
注意力机制
点监督机制
object detection
smart examination classroom
anchor-free method
attention mechanism
point supervision mechanism
