基于深度学习的油田在线视频目标检测  

Research and Application of Online Video Target Detection Based on Deep Learning

作　　者：张千[1] 梁鸿[1] 童彦淇 李洋[1] ZHANG Qian;LIANG Hong;TONG Yanqi;LI Yang(College of Computer Science and Technology,China University of Petroleum(East China),Qingdao 266580)

机构地区：[1]中国石油大学(华东)计算机科学与技术学院,青岛266580

出　　处：《计算机与数字工程》2024年第3期864-872,共9页Computer & Digital Engineering

摘　　要：油田背景复杂多变,摄像头悬挂较高,导致物体在监控画面中的比例较小,加大了检测难度。从油田实际场景出发,深入研究了SSD算法检测小目标准确率比较低的问题并对其改进,提出了RP-SSD算法,通过在特征金子塔中增加上采样模块和预测模块,更好地融合前后卷积层产生的特征图,并使用空洞卷积扩大了前面卷积层的感受野,提高了对小目标检测的准确率。采用Pascal VOC验证改进算法的有效性,同时选取了faster R-CNN、SSD300、DSSD321作为对照试验。试验结果表明,RP-SSD在小目标检测方面性能显著提高,可以达到实时检测的要求。The complex and ever-changing background of the oil field,coupled with high camera suspension,results in a smaller proportion of objects in the monitoring image,increasing the difficulty of detection.Starting from the actual oilfield scenar-io,the low accuracy of SSD algorithm in detecting small targets is deeply studied and improved.The RP-SSD algorithm is proposed by adding an upsampling module and a prediction module in the feature pyramid to better fuse the feature maps generated by the front and back convolutional layers.Hollow convolution is used to expand the receptive field of the front convolutional layers,im-proving the accuracy of small target detection.Pascal VOC is used to validate the effectiveness of the improved algorithm,and fast R-CNN,SSD300,and DSSD321 are selected as control experiments.The experimental results show that RP-SSD significantly im-proves its performance in small object detection and can meet the requirements of real-time detection.

关 键 词：小目标检测 特征金字塔 残差网络 空洞卷积 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]