针对遥感影像的MSA-YOLO储油罐目标检测

MSA-YOLO oil storage tank target detection for remote sensing images

原油作为一种重要的战略物资,在我国经济和军事等多个领域均起到重要作用。提出一种算法MSA-YOLO(MultiScale Adaptive YOLO),该算法在YOLOv4算法的基础上进行优化,并基于以吉林一号光学遥感卫星影像为主的遥感图像数据集进行实验,对特定监控区域内的储油罐进行识别与分类。算法优化内容包括:为简化储油罐监测模型同时保证模型的效率,对YOLOv4的网络结构中的多尺度识别模块进行修剪;使用k-means++聚类算法进行初始锚框的选取,使模型加速收敛;使用基于CIoU-NMS的优化,进一步提升推理速度和准确度。实验结果表明,与YOLOv4相比,MSA-YOLO模型参数数量减少25.84%;模型尺寸减少62.13%;在Tesla V100的GPU环境下,模型的训练速度提升6 s/epoch,推理速度提升15.76 F/s;平均精度为95.65%。与此同时,MSA-YOLO算法在多种通用目标识别算法进行的对比实验中均体现出了更高效的特点。MSA-YOLO算法对储油罐进行准确且实时的识别具有通用可行性,可为遥感数据在能源期货领域提供技术参考。

Crude oil,as an important strategic material,plays an important role in many fields such as my country’s economy and military.This paper proposes an algorithm MSA-YOLO(MultiScale Adaptive YOLO),which is optimized on the basis of the YOLOv4 algorithm,and is experimented based on the remote sensing image dataset mainly based on Jilin-1 optical remote sensing satellite images,to make identification and classification of oil storage tanks.The algorithm optimization contents include:in order to simplify the oil storage tank monitoring model and ensure the efficiency of the model,prune the multi-scale identification module in the network structure of YOLOv4;use the k-means++clustering algorithm to select the initial anchor frame to accelerate the convergence of the model;use CIoU-NMS-based optimization to further improve inference speed and accuracy.The experimental results show that compared with YOLOv4,the number of parameters of MSA-YOLO model is reduced by 25.84%;the model size is reduced by 62.13%;in the GPU environment of Tesla V100,the training speed of the model is increased by 6 s/epoch,and the inference speed is increased by 15.76 F/s;the average accuracy is 95.65%.At the same time,the MSA-YOLO algorithm shows more efficient characteristics in the comparative experiments of various general target recognition algorithms.The MSA-YOLO algorithm has universal feasibility for accurate and real-time identification of oil storage tanks,and can provide technical reference for remote sensing data in the field of energy futures.