摘要
针对遥感图像背景复杂、舰船目标尺度变化大和方向任意等问题,提出了一种基于YOLOv8的遥感图像舰船目标检测算法。首先为了提高特征融合的效率,在特征融合模块将原有的路径聚合网络优化为渐进式特征融合网络;其次为了强化模型的多尺度检测能力,在颈部网络加入基于大卷积核分解和空间选择机制的选择性大卷积核注意力模块;然后为了提高模型的分类和定位能力,在解耦检测头的检测框回归分支添加视觉注意力模块,在目标分类分支添加坐标注意力模块;最后为了将通用目标检测算法转换为旋转目标检测算法,在检测头引入角度参数并优化损失函数,优化网络训练时的正负样本标签分配策略以提升网络训练时的效率。在公开数据集HRSC2016上的实验结果表明,所提算法以3.02×10^(6)的参数量,实现了90.55%的检测准确率,与当前其他主流旋转目标检测算法相比,在参数量和准确率上都具有先进性。
Aiming at the problems of complex background of remote sensing images,large-scale changes and arbitrary directions of ship targets,a ship target detection algorithm for remote sensing images based on YOLOv8 is proposed.First,the original path aggregation network is optimized into an asymptotic feature fusion network in the feature fusion module to improve the efficiency of feature fusion.Second,a large selective kernel attention module based on large convolution kernel decomposition and spatial selection mechanism is added to the neck network,which strengthens the multi-scale detection ability of the model.Then the visual attention module is added to the detection frame regression branch and the coordinate attention module is added to the target classification branch of the decoupled detection head to improve the classification and positioning capabilities of the model.Finally,the angle parameter is introduced in the detection head and the loss function is optimized to convert the general object detection algorithm into a rotating object detection algorithm,and the label assignment strategy of positive and negative samples during network training is optimized to improve the training efficiency of the network.The experimental results on the public dataset HRSC2016 show that the proposed algorithm achieves a detection accuracy of 90.55%with a parameter quantity of 3.02×10^(6),which is advanced in terms of the number of parameters and accuracy compared with other current mainstream rotating object detection algorithms.
作者
姜忠旭
高磊
关智聪
辛苗
阮洋
JIANG Zhongxu;GAO Lei;GUAN Zhicong;XIN Miao;RUAN Yang(Shanghai Aerospace Control Technology Institute,Shanghai 201109;Infrared Detection Technology Research&Development Center of CASC,Shanghai 201109;Unit 93145 of the Chinese People s Liberation Army,Shanghai 200000)
出处
《飞控与探测》
2024年第3期56-66,共11页
Flight Control & Detection
关键词
舰船检测
遥感图像
旋转目标检测
图像处理
YOLO
ship detection
remote sensing images
rotating object detection
image processing
YOLO
