地理空间信息2024，Vol.22Issue(3)：24-28,5.
基于Att-DConv的遥感舰船检测方法研究
Research on Remote Sensing Ship Detection Method Based on Att-DConv
何民华 1张润达 1赵胜利2
作者信息
- 1. 交通运输部南海航海保障中心 广州海事测绘中心,广东 广州 510000
- 2. 广州市地铁设计研究院股份有限公司,广东 广州 510000
- 折叠
摘要
Abstract
Aiming at the problem of ship target detection in remote sensing images,we proposed a ship detection model based on deep learning.We used parallel dilated convolution groups and channel attention modules to form the backbone network at first.Then,we spliced the feature maps of different scales output by all feature extraction layers,and constructed four detection scales by upsampling and downsampling of the fused feature layers respectively.Finally,we adopted the improved NMS algorithm to optimize the final detection frame output.We used the the open source data set UCMerced_LandUse and FAIR1M combinative data set to train and test model,and used a variety of image enhancement algorithms to optimize the quality of training set.We used mosaic processing to obtain more training images with positive samples,performed tests on raw,unprocessed images.The results show that the average accuracy of model proposed in this paper can reach 0.91,the detection speed can reach 34 f/s,which has stable detection capability for ship samples with different complex backgrounds and scales.关键词
遥感影像/舰船检测/空洞卷积/通道注意力/融合特征增强Key words
remote sensing image/ship detection/dilated convolution/channel attention/fusion feature enhancement分类
天文与地球科学引用本文复制引用
何民华,张润达,赵胜利..基于Att-DConv的遥感舰船检测方法研究[J].地理空间信息,2024,22(3):24-28,5.
