徐慧芳,黄冬梅,贺琪,杜艳玲,覃学标,时帅,胡安铎.融合 scSE 模块的改进 Mask R-CNN海洋锋检测方法[J].海洋通报,2022,(1):
融合 scSE 模块的改进 Mask R-CNN海洋锋检测方法
Improved Mask R-CNN ocean front detection method fused with scSE module
投稿时间:2021-06-04  修订日期:2021-09-13
DOI:10.11840/j.issn.1001-6392.2022.01.003
中文关键词:  scSE 空间注意力  Mask R-CNN  海洋锋检测  Mask 损失函数
英文关键词:scSE spatial attention  Mask R-CNN  ocean front detection  Mask loss function
基金项目:国家自然科学基金青年基金 (41906179);上海市教育发展基金 (AASH2004);国家海洋局数字海洋科学技术重点实验室开放基金 (B201801029);上海市科委部分地方高校能力建设项目 (20050501900;20020500700)
作者单位E-mail
徐慧芳 上海海洋大学 信息学院上海 201306上海建桥学院 信息技术学院上海 201306 17069@gench.edu.cn 
黄冬梅 上海海洋大学 信息学院上海 201306上海电力大学上海 201306上海电力大学 国际交流与合作处上海 201306 17069@gench.edu.cn 
贺琪 上海海洋大学 信息学院上海 201306  
杜艳玲 上海海洋大学 信息学院上海 201306  
覃学标 上海海洋大学 信息学院上海 201306  
时帅 上海电力大学 国际交流与合作处上海 201306  
胡安铎 上海电力大学上海 201306  
摘要点击次数: 194
全文下载次数: 33
中文摘要:
      海洋锋是重要的中尺度海洋现象,具有数据量小、目标小、弱边缘等特性。针对实际检测任务中弱边缘、小目标海洋锋的检测精度低、错检及漏检率高等问题,融合 scSE (spatial and channel Squeeze & Excitation) 空间注意力模块构建了一种改进的 Mask R-CNN 海洋锋检测模型。该方法首先对 Mask R-CNN 骨干网络结构进行改进,采用 scSE 模块引导的ResNet-50 网络作为特征提取网络,通过加权策略对图像通道和空间位置进行特征突出,提升网络对重要特征的提取能力;其次,针对海洋锋目标边缘定位不准确的问题,引入 IoU boundary loss 构建新的 Mask 损失函数,提高边界检测精度。最后,为验证方法的有效性,从训练数据和实验模型上,分别设计多组对比实验。实验结果表明,相比传统 Mask R-CNN、 YOLOv3 神经网络及现有 Mask R-CNN 改进网络,本文方法对 SST 梯度影像数据集上的强、弱海洋锋检测效果最好,定位准确率 (IoU,Intersection-over-union)) 及检测精度 (mAP,Mean Average Precision) 均达 0.914 以上。此外,对文中设计评 估模型进行检测效率实验,结果发现在不同网络模型、不同迭代次数情况下,本文提出模型消耗时间最短,远低于 YOLOv3网络完成训练时所用时长。
英文摘要:
      Ocean front is an important mesoscale ocean phenomenon, with the characteristics of small data volume, small targets, and weak edges. To solve the problems of low detection accuracy, high false and missing detection on weak edges and small target ocean fronts in actual detection tasks, this paper constructed an improved Mask R-CNN ocean front detection model by fusing SCSE (spatial and channel squeeze & exception) spatial attention module. Firstly, the structure of Mask R-CNN backbone network is improved, and the resnet-50 Network Guided by SCSE module is used as the feature extraction network, and the weighting strategy is used to highlight the image channels and spatial locations to improve the ability of network to extract important features. Secondly, in response to the problem of inaccurate positioning of the edge of the target ocean front, constructing a new mask loss function is by introducing IoU boundary loss to improve the accuracy of boundary detection. Finally, in order to verify the effectiveness of the method, several groups of comparative experiments are designed from the training data and the experimental model. The experimental results show that compared with the traditional Mask RCNN, YOLOv3 and the existing improved Mask R-CNN network, the method in this paper has the best effect on strong and weak ocean front detection on the SST gradient image data set, the accuracy of positioning (IoU) and detection accuracy (mAP) are more than 0.914. In addition, the detection efficiency experiment is designed in this paper. The results show that themethod in this paper takes the shortest time under different network models and different iteration times, which is much lower than the YOLOv3 network.
查看全文    下载PDF阅读器
关闭