基于ATD-CNN模型的黄河郑州段水面漂浮物检测研究  

Research on the Detection of Floating Objects on the Water Surface of the Zhengzhou Section of the Yellow River Based on the ATD-CNN Model

作　　者：邵晓艳[1] 王军 赵雪专 王胜 冯军 SHAO Xiaoyan;WANG Jun;ZHAO Xuezhuan;WANG Sheng;FENG Jun(School of Computer Science,Zhengzhou University of Aeronautics,Zhengzhou 450046,China;School of Computer and Information Engineering,Henan University,Kaifeng 475004,China)

机构地区：[1]郑州航空工业管理学院计算机学院,河南郑州450046 [2]河南大学计算机与信息工程学院,河南开封475004

出　　处：《人民黄河》2025年第2期131-136,共6页Yellow River

基　　金：国家自然科学基金资助项目(U1904119);河南省科技攻关计划项目(232102210033,232102210054);河南省重点研发专项(231111212000);河南省杰出外籍科学家工作室项目(GZS2022011);航空科学基金资助项目(20230001055002);重庆市自然科学基金资助项目(CSTB2023NSCQ-MSX0070)。

摘　　要：针对水面漂浮物感知目标小、易受干扰、识别精度低的问题,提出ATD-CNN目标检测模型。结合注意力机制,将注意力模块嵌入Faster R-CNN改进模型的基本主干网络,计算特征图内部特征点之间的长距离相关系数,对显著性特征进行有效增强,以提升基本主干网络对图像特征的提取能力。基于河南省郑州市惠济区南裹头黄河沿岸采集的图像数据,对ATD-CNN模型检测效果进行验证,并将该模型性能与Faster R-CNN改进模型、YOLOv5单阶段目标检测模型进行对比。结果表明:与Faster R-CNN改进模型相比,ATD-CNN模型对水面漂浮物的漏检率下降,其mAP值提升了6.80%,F1 Score平均值提升了2%。与YOLOv5X、Faster R-CNN改进模型相比,ATD-CNN模型的mAP值分别提升了2.91%、6.80%,有效提高了水面漂浮物检测精度。Aiming at the issues of small targets,vulnerable to interference and low recognition accuracy of floating objects on the water surface,ATD-CNN object detection model was proposed.Combined with the attention mechanism,the attention module was embedded into the basic backbone network of the Faster R-CNN improved model,and the long-distance correlation coefficient between the feature points in the feature map was calculated to effectively enhance the saliency features,so as to improve the ability of the basic backbone network to extract image features.Based on the image data collected along the Yellow River in Nanbaotou,Huiji District,Zhengzhou City,Henan Province,the detection effectiveness of the ATD-CNN model was verified,and the performance of the model was compared with the Faster R-CNN improved model and YOLOv5 single-stage object detection model.The results show that comparing with the Faster R-CNN improved model,the ATD-CNN model reduces the missed detection rate of floating debris on the water surface,increases its mAP value by 6.80%,and increases the average F1 Score by 2%.Comparing with YOLOv5X and Faster R-CNN improved models,the mAP values of ATD-CNN model increase by 2.91%and 6.80%respectively,effectively improving the accuracy of floating object detection on the water surface.

关 键 词：卷积神经网络 水面漂浮物 目标检测 注意力 黄河郑州段 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TV882.1[自动化与计算机技术—计算机科学与技术]