基于残差网络和改进特征金字塔的油田作业现场目标检测算法

针对单点多盒检测器(single shot multibox detector,SSD)对小目标识别率低的问题,提出一种基于残差网络和改进特征金字塔(feature pyramid networks,FPN)的RP-SSD(residual and pyramid SSD)算法,并将其应用于油田安防领域。为了得到小物体更多的信息,首先在特征金字塔中增加上采样模块,并在上采样模块之后添加预测模块,之后采用空洞卷积增大Conv4_3的感受野。RP-SSD网络变深,针对RP-SSD在反向传播过程中存在梯度爆炸或梯度消失的问题,采用跳层连接的方式改进基础网络。RP-SSD在PASCAL VOC测试的准确率(meanaverage precision,mAP)为78.9%,比SSD提高了1.7%,其中对于目标较小的bottle类提高了8.9%。实验结果表明,RP-SSD对小目标检测的性能提高显著,同时RP-SSD在GTX 1080Ti上测试的速度为32帧/s,可见RP-SSD可以达到实时处理的要求。

Field Target Detection Algorithms for Oilfield Operation Based on Residual Network and Improved Feature Pyramid Networks

To solve the problem of low recognition rate of small targets by single shot multibox detector(SSD), a RP-SSD(residual and pyramid SSD) algorithm based on residual network and improved feature pyramid networks is proposed, and RP-SSD is applied to the field of oil field security. In order to get more information about small objects, the up sampling module is added to the feature pyramid networks first, and the prediction module is added after the up sampling module. Then, the receptive field of Conv4_3 is increased by using dilated convolution. RP-SSD network deepens. In order to solve the problem of gradient explosion or Gradient disappearance in the back propagation process of RP-SSD, the basic network is improved by means of Jump-Layer connection. The mAP (mean average precision) of RP-SSD in PASCAL VOC test was 78.9%. RP-SSD was 1.7% higher than mAP of SSD, and 8.9% higher for bottle class with smaller target. The experimental results show that the performance of RP-SSD for small target detection is significantly improved. At the same time, the speed of RP-SSD testing on GTX 1080Ti is 32 fps (frame per second). It can be seen that RP-SSD can meet the requirements of real-time processing.