摘要
针对复杂场景下合成孔径雷达(synthetic aperture radar,SAR)图像目标检测的鲁棒性差和准确性低等问题,提出一种基于感兴趣区域池化的方法进行SAR图像目标检测。在YOLOv3模型中,引入可变形卷积,加入池化层,通过图像增强的方式对感兴趣区域的目标进行特征选择;通过增加全连接层,生成每个位置的偏移量进行扩张,在偏移量中增加权值;在FPN部分增加DropBlock模块;改进YOLOV3的训练策略方法,采用平滑地调整学习率和增量的方式进行训练,提升模型对SAR图像感兴趣目标的检测准确率。在SAR图像上进行验证,精度可以达到98.2%,验证了模型的有效性。
Aiming at the problems of the poor robustness and accuracy of the synthetic aperture radar(SAR)image target detection in complex scenes,a method based on the pooling of regions of interest for SAR image target detection was proposed.In YOLOv3 model,deformable convolution was introduced,pooling layer was added,and features of targets in regions of interest were selected through image enhancement;by adding the full connection layer,the offset of each position was generated for expansion,and the weight value was added to the offset;adding DropBlock module in FPN;the training strategy and method of YOLOV3 were improved.The training was carried out in a smooth and incremental way to adjust the learning rate,so as to improve the detection accuracy of the model for the target of interest in SAR images.The accuracy of SAR image is 98.2%,which proves the validity of the model.
作者
郭瑞香
GUO Ruixiang(Information Construction and Management Office,Minnan Normal University,Zhangzhou 363000,China)
出处
《邵阳学院学报（自然科学版）》
2023年第2期29-36,共8页
Journal of Shaoyang University：Natural Science Edition
基金
福建省高校教育信息化科研课题(FJGX22005)。
