摘要
在实际陆战场环境中,常常因为所检测的实际目标(如坦克、火炮、单兵等)存在多尺度识别率低、正负样本不均衡的难题,导致无法准确实时地实现陆战场目标检测。虽然现有的One-stage检测器提高了一定的速率但是不能满足准确率的要求。为此提出一种基于改进的多层次特征金字塔网络(I-MLFPN)的陆战场目标检测算法,在提高识别精度的同时提升检测速度。以目标检测框架SSD为基础,提出改进的多层次特征金字塔网络(I-MLFPN)来构建更有效的特征金字塔,用于检测不同尺度的对象;采用Focal loss损失函数,解决因正负样本不均衡带来低准确率的问题。该算法将目标检测速度提高到24.8帧/s,目标检测准确率提高了7%,达到69.3%,有效解决了检测速度与准确率兼得的问题。
In the actual land battlefield environment,the problems of low multi-scale recognition rate and imbalance between positive and negative samples are often accounted in the actual targets detections(such as tanks,artillery,individual soldiers and other equipment),which leads to the inaccuracy of real-time target detection in the land battlefield.Although the existing One-stage detector improves the speed,it can not meet the requirements of accuracy.Therefore,a land battlefield target detection algorithm based on improved multi-level feature pyramid network(I-MLFPN)is proposed to improve the recognition accuracy and detection speed.On the basis of target detection framework SSD,an improved multi-level feature pyramid network(I-MLFPN)is proposed to construct a more effective feature pyramid for detecting objects of different scales.Focal loss function is applied to solve the problem of low accuracy caused by the imbalance of positive and negative samples.The algorithm improves the speed of target detection to 24.8 Frames/s,and the target detection accuracy is increased by 7%to 69.3%.It effectively solves the problem of both detection speed and accuracy.
作者
吴娇
王鹏
乔梦雨
贺咪咪
Wu Jiao;Wang Peng;Qiao Mengyu;He Mimi(College of Electronics and Information Engineering,Xi’an Technological University,Xi’an 710021,Shaanxi,China)
出处
《计算机应用与软件》
北大核心
2020年第10期155-161,共7页
Computer Applications and Software
基金
国家自然科学基金项目(61671362)
陕西省科技厅重点研发计划项目(2019GY-022)。
关键词
深度学习
多层次特征金字塔
目标检测
通道混洗
Deep learning
Multi-level feature pyramid
Target detection
Channel shuffle
