摘要
电力设备的锈蚀检测作为危害电力系统安全运行的重要一环必须能被快速、准确地进行识别与检测并及时报警。为了提高电力设备锈蚀区域检测的时效性和可靠性,基于YOLOv3目标检测算法,结合注意力机制提出了一种改进的Attention-YOLOv3算法,可以实现对锈蚀区域的快速可靠识别。首先,利用深度可分离卷积对YOLOv3的特征提干网络进行轻量化处理来缩减模型的大小,提高检测速度。其次,为了弥补轻量化网络带来的精度损失,提高特征的提取能力,在上采样之后采用了空间注意力机制(spatial-attention)和通道注意力机制(channel-attention)结合的级联双注意力机制对特征进行融合筛选,剔除冗余的无效特征。实验表明,提出的锈蚀区域检测算法能有效地检测和识别出电力设备的锈蚀区域,相比较标准YOLOv3可以做到在检测时间缩短近46%的情况下提升9.06%的检测精度,在RustDetection数据集上可以达到91.75%的平均精度。
Corrosion detection of power equipment,as an important part of the safety operation of the power system,must be quickly and accurately identified and detected and timely alarmed.In order to improve the timeliness and reliability of corrosion area detection of power equipment,we propose an improved Attention-YOLOv3 algorithm based on YOLOv3 target detection algorithm combined with attention mechanism,which can realize fast and reliable identification of rust area.Firstly,we reduce the model size and improve the detection speed by lightweighting the YOLOv3 backbone network with the depthwise convolution.Then,in order to compensate for the loss of precision caused by lightweight network and improve the feature extraction ability,we use a cascade of double-spatial-channel(channel-attention)and channel-attention(spatial-attention)after upsample operation.The attention mechanism combines and filters the features to eliminate redundant invalid features.Experiments show that the proposed rust area detection algorithm can effectively detect and identify the rust area of power equipment.Compared with the standard YOLOv3,the detection accuracy can be improved by 9.06%when the detection time is shortened by nearly 46%.The proposed algorithm can achieve an average precision of 91.75%at the RustDetection dataset.
作者
吴之昊
熊卫华
任嘉锋
姜明
WU Zhi-hao;XIONG Wei-hua;REN Jia-feng;JIANG Ming(Faculty of Mechanical Engineering&Automation,Zhejiang Sci-Tech University,Hangzhou 310018,China;School of Computer Science and Technology,Hangzhou Dianzi University,Hangzhou 310018,China)
出处
《计算机技术与发展》
2020年第11期147-152,共6页
Computer Technology and Development
基金
国家自然科学基金(61803339,61503341)
浙江省自然科学基金(LQ18F030011)
浙江省重点研发计划项目(2019C03096)。
关键词
目标检测
锈蚀检测
注意力机制
特征提取网络
轻量级网络
object detection
corrosion detection
attention mechanism
feature extraction network
lightweight network