基于改进目标检测模型的油田配注站场仪表盘检测方法

PDF下载HTML

Title:

Instrument panel detection method for oilfield injection station based on improved target detection model

作者:

张强¹, 何子成¹, 盛守东²

1. 东北石油大学 计算机与信息技术学院，黑龙江 大庆 163318;
2. 大庆油田有限责任公司第五采油厂 地质研究所，黑龙江 大庆 163513

Author(s):

ZHANG Qiang¹, HE Zicheng¹, SHENG Shoudong²

1. College of Computer and Information Technology, Northeast Petroleum University, Daqing 163318, China;
2. Institute of Geology, The Fifth Oil Production Plant of Daqing Oilfield Co., Ltd., Daqing 163513, China

关键词:

小目标检测; Faster R-CNN; 双分支融合; 移动卷积网络; 注意力机制; 轻量化网络; 夜间低光照检测; 边缘设备部署

Keywords:

small object detection; Faster R-CNN; dual branch fusion; mobile convolutional network; attention mechanism; lightweight network; low-light detection; edge device deployment

分类号:

TP391

DOI:

10.11992/tis.202508036

摘要:

在油田配注站场的夜间作业环境中，仪表盘目标检测受限于低光照条件，导致传统单模态图像检测方法在准确性与鲁棒性方面表现不佳。本文提出了一种改进快速区域卷积神经网络(faster region-based convolutional neural networks, Faster R-CNN)的中期融合双分支目标检测模型，旨在融合可见光图像与夜视图像的互补信息，提升夜间低对比度环境下的目标检测性能。设计针对可见光与夜视图像的双分支中期融合策略，增强双图像特征协同，通过改进轻量化的移动卷积网络为主干特征提取网络，结合空间注意力机制有效增强了关键区域的特征表达能力。在自定义油田配注站场仪表盘夜间图像数据集上的实验结果表明，该方法在mAP@0.5和mAP@0.5:0.95指标上分别达到了88.92%与46.87%，较基线Faster R-CNN模型分别提升了5.89和4.72百分点，在低对比度的夜间工业场景中展现出良好的检测精度与鲁棒性，同时降低模型的参数量，使其更容易部署在资源有限的边缘设备。

Abstract:

In the night operation environment of oilfield distribution stations, instrument panel target detection is limited by low light conditions, resulting in poor performance of traditional single-modal image detection methods in terms of accuracy and robustness. This paper proposes a mid-term fusion dual-branch target detection model based on improved Faster R-CNN, which aims to fuse the complementary information of visible light images and night vision images to improve the target detection performance in low-contrast environments at night. A dual-branch mid-term fusion strategy for visible and night vision images is designed to enhance the synergy of dual image features. By improving the lightweight mobile convolutional network as the backbone feature extraction network, the feature expression ability of key areas is effectively enhanced by combining the spatial attention mechanism. Experimental results on a custom oilfield distribution station instrument panel night image dataset show that this method achieves 88.92% and 46.87% in mAP@0.5 and mAP@0.5:0.95 indicators, respectively, which are 5.89% and 4.72% higher than the baseline Faster R-CNN model, respectively. It shows good detection accuracy and robustness in low-contrast night industrial scenes, while reducing the number of model parameters, making it easier to deploy on resource-limited edge devices.