摘要
针对人工设计特征复杂和传统的自动化方法在目标识别和检测上准确率和效率低下的问题,提出一种基于改进的Faster-RCNN的目标区域定位方法。由于被检测目标尺度变化大,为了解决原始Faster-RCNN网络结构对于小目标区域检测精度低的问题,提出多尺度快速区域卷积神经网络检测算法,改进了神经网络的结构,使网络在检测过程中可以同时使用低层和高层的特征,提升了网络对于小目标区域的检测能力。修改原始网络中锚框设定方法,通过聚类算法来确定不同尺度的特征图的锚框。实验结果表明,该方法在不同的背景下均能实现对目标较好的识别与定位,对小区域的检测能力显著提高。检测精度由原始网络结构的79.60%上升到95.39%,提高了15.79百分点。
Aiming at the complexity of artificial design features and the low accuracy and efficiency of traditional automatic methods in target recognition and detection,this paper proposes a method for vehicle paint defects area location based on improved Faster-RCNN.Due to significant changes in the scale of the detected target,in order to solve the problem that the original Faster-RCNN network structure had low detection accuracy for small area,a multi-scale fast area convolution neural network detection algorithm was proposed.The structure of neural network was improved,so that the network could use the feature of low-level and high-level during the detection process at the same time,which improved the detection ability of the network for small area.The anchor set in the original network was modified,and the anchor of different scale feature map was settled by clustering algorithm.The experimental results show that the proposed method can achieve better target recognition and location under different background,especially the detection ability of small area is greatly improved,and the detection accuracy is increased from 79.60%of the original network structure to 95.39%and increased by 15.79%.
作者
薛阳
叶晓康
孙越
洪俊
万轶伦
Xue Yang;Ye Xiaokang;Sun Yue;Hong Jun;Wan Yilun(School of Automation Engineering,Shanghai Electric Power University,Shanghai 200090,China;State Grid Shanghai Pudong Power Supply Company,Shanghai 200122,China)
出处
《计算机应用与软件》
北大核心
2023年第8期193-200,共8页
Computer Applications and Software
基金
国网浙江省电力有限公司科技项目(5211HZ17000F)
上海市电站自动化技术重点实验室(13DZ2273800)。
关键词
目标检测与定位
深度学习
多尺度检测
聚类算法
锚框
Target detection and location
Deep learning
Multiscale detection
Clustering algorithm
Anchors
