基于改进的YOLOv7油田井场压力表小目标识别

Identification of Small Target Pressure Gauges in Oilfield Well Sites Based on Improved YOLOv7

针对油田井场压力表小目标在不同光照条件下特征提取困难且目标检测率低等问题,提出一种基于改进的YOLOv7小目标井场压力表实时检测算法。首先,利用全局最大池化和全局最小池化的简单线性组合生成一种自适应池化层,并使用轻量级模块生成SPEM注意力模块,后将其加入YOLOv7主干网络中,提高网络对不同光照条件下压力表小目标的特征提取;其次,使用模量激活函数替代SiLU函数,减少死亡节点,解决梯度消失问题,提高模型的泛化能力;最后,将原来的YOLOv7中的CIoU损失函数替换为Wise-IoU,实现对损失函数的优化,并利用梯度增益分配策略来减少低质量示例产生的有害梯度,以聚焦普通质量锚盒的预测和回归。实验结果表明,改进后的算法相较于原YOLOv7算法精准率上升1.07%,召回率提升2.08%,同时也优于Faster RCNN、SSD、YOLOv3算法的检测结果,能有效满足油田井场压力表小目标的检测要求,具有较强的工程实践意义。

In order to solve the problems of difficult feature extraction and low detection rate of small target pressure gauges on oilfield well site under different lighting conditions,a real-time well site small target pressure gauge detection algorithm based on an improved YOLOv7 is proposed.Firstly,a simple linear combination of global max pooling and global min pooling is used to generate an adaptive pooling layer,and a lightweight module is used to generate an SPEM attention module,which is added to the YOLOv7 backbone network to improve the network's feature extraction of small target pressure gauges under different lighting conditions.Secondly,SiLU function is replaced by modulus activation function to reduce dead nodes,solve the problem of gradient vanishing,and improve the model's generalization ability.Finally,the CIoU loss function in the original YOLOv7 is replaced with Wise-IoU to optimize the loss function,and the harmful gradients generated by low-quality examples is reduced by using a gradient gain allocation strategy to focus on the prediction and regression of ordinary quality anchor boxes.The experimental results show that compared to the original YOLOv7 algorithm,the improved algorithm improves accuracy by 1.07%and recall rate by 2.08%.At the same time,it also outperforms the detection results of Faster RCNN,SSD,and YOLOv3 algorithms,and can effectively meet the detection requirements of small target pressure gauges in oilfield well sites,with strong engineering practical significance.