基于UPCBAM-RYOLO V5 的光学遥感舰船小目标检测

吉林大学学报(信息科学版) ›› 2024, Vol. 42 ›› Issue (6): 1048-1057.

基于UPCBAM-RYOLO V5 的光学遥感舰船小目标检测

杨笑天^1,2, 鱼昕^1,2, 刘铭³, 王梁^1,2, 谭金林^1,2, 吴意^1,2

1. 西安空间无线电技术研究所,西安710100;2. 陕西航天技术应用研究院有限公司系统总体部, 西安710100; 3. 长春工业大学数学与统计学院, 长春130012

收稿日期:2023-11-07 出版日期:2024-12-23 发布日期:2024-12-23
通讯作者: 鱼昕(1996— ), 男, 陕西富平人, 西安空间无线电技术研究所工程师,主要从事遥感图像智能解译、多模态智能模型研究,(Tel)86-17808098687(E-mail)2252685386@qq.com
作者简介:杨笑天(1991— ), 男, 宁夏海原人, 西安空间无线电技术研究所工程师, 硕士, 主要从事人工智能、遥感图像处理技术研究,(Tel)86-18700948001(E-mail)415542866@ qq. com
基金资助:
吉林省科技厅基金资助项目(20200201157JC)

Optical Remote Sensing Ship Small Target Detection Based on UPCBAM-RYOLO V5

YANG Xiaotian^1,2, YU Xin^1,2, LIU Ming³, WANG Liang^1,2, TAN Jinlin^1,2, WU Yi^1,2

1. China Academy of Space Technology(Xi’an), Xi’an 710100, China; 2. General Department of the System, Shaanxi Aerospace Technology Application Research Institute Company Limited, Xi’an 710100, China; 3. School of Mathematics and Statistics, Changchun University of Technology, Changchun 130012, China

Received:2023-11-07 Online:2024-12-23 Published:2024-12-23

摘要/Abstract

摘要： 针对光学遥感数据中舰船小目标数量占比大,长宽比大且多个目标紧密排列难以检测的问题,提出一种基于YOLO V5(You Only Look Once V5)的光学遥感舰船小目标检测算法———UPCBAM-RYOLO V5(Upsampling Convolutional Attention Block Module-RYOLO V5)算法。该算法设计了一种上采样注意力机制模块, 增强了网络对小尺寸目标特征提取能力,并在边框回归公式中引入旋转角度损失,提高了算法对舰船外观和方向的感知能力。基于GF1、GF2组成的舰船小目标数据集实验, 结果表明, UPCBAM-RYOLO V5 算法模型提升了舰船小目标检测的定位精度和分类精度,其中P、R、 MAP(Mean Average Precision)值分别达到93%、94%和95%, 较传统YOLO V5 模型分别提高3%、7%和7%。对网络中上采样注意力机制模块添加位置的消融实验结果表明, 相较于在Backbone和Prediction 处加入 UPCBAM, 在 Neck 处加入 UPCBAM 对遥感影像舰船小目标的检测影响最大,性能最优,P、R和MAP值分别提高了5%、4%和2%。从而验证了UPCBAM-RYOLO V5模型在光学遥感舰船小目标检测方面具有一定的研究意义。

关键词: 光学遥感数据, 舰船小目标, UPCBAM-RYOLO V5算法, 上采样注意力机制, 旋转角度

Abstract: In order to solve the problems of large proportion of small targets in optical remote sensing data, the aspect ratio is large, and multiple targets are closely arranged and difficult to detect,we present an optical remote sensing small ship target detection algorithm based on the YOLO V5(You Only Look Once V5), UPCBAM- RYOLO V5 (Upsampling Convolutional Attention Block Module-RYOLO V5) algorithm. An up-sampling attention mechanism module is designed to enhance the feature extraction ability of small size targets. The rotation angle loss is introduced into the frame regression formula to improve the algorithm’s perception ability of the ship’s appearance and direction. Based on the experiment of small ship target datasets composed of GF1 and GF2, the results show that the UPCBAM-RYOLO V5 algorithm model improves the positioning accuracy and classification accuracy of small ship target detection. The P value, R value, and MAP(Mean Average Precision) value reach 93%, 94%, and 95% respectively, which are 3%, 7%, and 7% higher than the original YOLO V5 model. In the upsampled attention-mechanism module added location ablation experiment to the network, the results show that compared with the addition of UPCBAM in Backbone and Prediction, the addition of UPCBAM in Neck has the greatest influence on the detection of small target ships in remote sensing images. The performance is the best, with P value, R value, and MAP value increased by 5%, 4%, and 2%, respectively. The UPCBAM-RYOLO V5 model is proved to have a certain research significance in optical remote sensing ship small target detection.

Key words: optical remote sensing data, small ship targets, upsampling convolutional attention block module-R you only look once V5(UPCBAM-RYOLOV5) algrithom, upsampling convolutional attention block module, rotation angle

中图分类号:

TP319.4

杨笑天, 鱼昕, 刘铭, 王梁, 谭金林, 吴意. 基于UPCBAM-RYOLO V5 的光学遥感舰船小目标检测[J]. 吉林大学学报(信息科学版), 2024, 42(6): 1048-1057.

YANG Xiaotian, YU Xin, LIU Ming, WANG Liang, TAN Jinlin, WU Yi. Optical Remote Sensing Ship Small Target Detection Based on UPCBAM-RYOLO V5[J]. Journal of Jilin University (Information Science Edition), 2024, 42(6): 1048-1057.