基于特征融合及混合注意力的小目标船舶识别
Research on small target ship recognition based on feature fusion method and hybrid attention model
严荣慧 1郭前 1雷鸣 1蔡雁翔 2羊箭锋2
作者信息
- 1. 苏州城市学院 智能制造与智慧交通学院,江苏 苏州 215006
- 2. 苏州大学 电子信息学院,江苏 苏州 215006
- 折叠
摘要
[目的]旨在解决由于小目标船舶特征不显著导致网络模型识别率低的问题,提出基于图像与运动特征融合方法.[方法]在小目标船舶图像特征不明显的情况下,融合运动特征信息,丰富小型船舶的特征表达,同时提出混合注意力模型.在数据驱动条件下加入船舶目标先验信息,提高模型对关键特征的感知和利用能力.[结果]在 720P分辨率图像中能实现 10×10 像素的小目标船舶识别,识别距离达到 4 km范围,实现了广域船舶识别与定位功能.[结论]改进后的目标识别网络既具备像素级别的小目标船舶检测能力,同时又具备对环境噪声的抗干扰能力,突破了网络模型对小目标船舶识别率低的瓶颈.
Abstract
[Objective]This paper addresses a problem affecting small target ship detection in which net-work models have a low recognition rate caused by the insignificance of features in small target ships.[Methods]A fusion method based on the integration of image and motion features is proposed to enrich the feature representation of small ships in scenarios where the features of small target ship images are not promin-ent.Additionally,a hybrid attention model incorporates the prior information of ship targets under data-driven conditions to enhance the model's perception and utilization of key features.[Results]The proposed meth-od achieves the recognition of small target ships with a resolution of 720P at a distance of up to 4 kilometers,enabling wide-area ship recognition and localization functionality.[Conclusion]The improved target recog-nition network exhibits pixel-level small target detection capability while also demonstrating robustness against environmental noise interference,thereby overcoming the bottleneck of the low recognition rate of net-work models in small target ship detection.
关键词
目标跟踪/图像处理/小目标船舶/运动特征/特征融合/先验信息/混合注意力模型/YOLOv5Key words
target tracking/image processing/small target ships/motion features/feature fusion/prior information/hybrid attention model/YOLOv5引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据