Re-YOLOX:利用Resizer改进的YOLOX近岸海域监测目标识别模型  

Re-YOLOX:A YOLOX model for identifying nearshore monitoring targets improved based on the Resizer model

作　　者：王振华[1,2] 谭智联 李静 常英立 WANG Zhenhua;TAN Zhilian;LI Jing;CHANG Yingli(College of Information Technology,Shanghai Ocean University,Shanghai 201306,China;Key Laboratory of Marine Environmental Survey Technology and Application,Ministry of Natural Resources,Guangzhou 510000,China)

机构地区：[1]上海海洋大学信息学院,上海201306 [2]自然资源部海洋环境探测技术与应用重点实验室,广州510000

出　　处：《自然资源遥感》2023年第3期10-16,共7页Remote Sensing for Natural Resources

基　　金：自然资源部海洋环境探测技术与应用重点实验室开放基金项目“基于深度学习的海岛与海岸带典型要素智能监测关键技术研究与试点应用”(编号:MESTA-2021-B007);上海市地方院校能力建设项目“复杂潮汐环境影响下海岛(礁)地物信息提取与精度验证方法及其示范应用”(编号:19050502100)共同资助。

摘　　要：近岸海域监测包括自然环境监测和人类活动监测,其监测目标的高精准识别对海洋经济的健康发展、海洋环境的生态保护以及海洋防灾减灾等都有重要的作用。近岸海域监测目标具有多类型、多尺寸和不确定性等特征,现有识别模型在对近岸海域监测目标识别时,存在精度和效率欠佳、小目标漏检现象严重等问题。针对上述问题,利用可学习的图像调整模型(Resizer model)改进YOLOX,提出了面向近岸海域监测目标的识别模型(Re-YOLOX),包括:①利用Resizer model加强模型训练,提升模型的特征学习能力和表达能力,提高模型的召回率;②改进YOLOX的特征金字塔融合结构,减少小目标识别的漏检问题。用无人机监测的近岸海域视频数据作数据集,以车辆、船只和堆砌物为监测目标,将提出的Re-YOLOX模型与CenterNet,Faster R-CNN,YOLOv3和YOLOX等模型进行比较。结果表明,Re-YOLOX模型的平均预测精准率mAP可达94.23%,平均召回率mR可达91.99%,平均F1值mF1可达89.67%,均高于对比模型。综上所述,文章提出Re-YOLOX在保证目标识别效率的前提下提高了目标识别的精度,可为近岸海域管理提供技术支撑。Nearshore monitoring covers natural environments and human activities.High-accuracy identification of nearshore monitoring targets significantly influences the healthy development of the marine economy,the ecological protection of marine environments,and the prevention and mitigation of marine disasters.The nearshore monitoring targets feature multiple types,diverse sizes,and uncertainty.The existing identification models suffer low accuracy,low efficiency,and severe omission of small targets.This study proposed an identification model(Re-YOLOX)for nearshore monitoring targets by improving YOLOX using a learnable image resizer model(the Resizer model).First,the model training was intensified using the Resizer model to improve the feature learning and expression abilities and the recall rate of the Re-YOLOX model.Then,the feature pyramid fusion structure of the YOLOX algorithm was improved to reduce the omission of small targets in the identification.With the nearshore video data from UAV monitoring as the data set and cars,ships,and piles as monitoring targets,this study compared the Re-YOLOX model with other models,including CenterNet,Faster R-CNN,YOLOv3,and YOLOX.The results show that the Re-YOLOX model yielded a mean average precision of 94.23%,a mean recall of 91.99%,and a mean F1 score of 89.67%,all of which were higher than those of the other models.In summary,the Re-YOLOX model can improve the target identification accuracy while ensuring target identification efficiency,thus providing technical support for managing nearshore seas.

关 键 词：近岸海域 目标识别 YOLOX算法 无人机监测数据 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置] TP183[自动化与计算机技术—控制科学与工程]