联合双任务和双重注意力机制的遥感影像舰船提取方法

打开文本图片集

关键词：遥感影像；舰船提取；数据集；注意力机制；双任务中图分类号：TP391.41 文献标识码：Adoi：10.37188/OPE.20253314.2303 CSTR：32169.14.OPE.20253314.2303

Joint dual-task and dual-attention mechanism for ship extraction from remote sensing images

ZOU Yujie，LIN Yuzhun，WANG Shuxiang，ZUO Xibing，LIU Xiao，RUI Jie* （information engineer university. Zhengzhou 45OoOl，China） * Corresponding author，E-mail： zwjllhtt@scut.edu. cn

Abstract： Aiming at the problems of insuficient data baseline and poor decoding accuracy in the pixel-lev el ship-extraction task of remote-sensing images，this paper conducted research from the two levels of dataset and network structure，constructed a Ship Semantic Segmentation Dataset （SSSD） that was sufficient in quantity，rich in type，and wide in distribution，and proposed a ship-extraction method with joint dualtask and dual-attention mechanisms.At the dataset level，port areas worldwide were reasonably selected， the construction of SSSD was completed on the basis of the HRSC2Ol6 dataset，and the final dataset contained 3 76O training images and 923 test images. At the network-structure level， segmentation accuracy was effectively improved by fusing multi-head self-attention and channel attention and by designing the dual-attention mechanism.In addition，contour-supervised branching was introduced as a secondary task， and the dual-task network framework was formed by jointly performing semantic segmentation，which enhanced the separation between the ship and the background. The experimental results show that this paper's method improves 11. 44% ， 17.84% ， 12.25% ，and 1.64% over U-Net，ASPPUNet，DeepLabV3+ ，and MANet， respectively，on IoU； the design of dual-attention and auxiliary tasks improves 2.02% and 1. 36% ，respectively，on IoU；meanwhile，this paper's method exhibits，in generalization experiments，a stronger adaptabilityand robustness.The SSSD constructed in this paper can provide a reliable data benchmark for ship pixel-level decoding，and the ship extraction method proposed in this paper has better performance as wellas stronger stability and generalization capability compared with mainstream methods.

Key words： remote sensing image；ship extraction； data set；attention mechanism；dual task

1引言

我国拥有广阔的海洋领土和丰富的海洋资源，舰船作为维护海洋权益的核心作战力量和海洋侦察的主要目标，历来备受关注。（剩余17803字）