| 摘要: |
| 高分辨率合成孔径雷达（Synthetic Aperture Radar，SAR）图像中不同目标的尺寸区别较大，这使得小目标的特征不明显，为目标检测带来了极大的挑战。针对这一问题，提出了SAR YOLO 960算法。该算法首先改进了图像输入大小的限制，将输入图像提升到960×960像素；进而改善了YOLOv3(You Only Look Once v3)网络的整体结构，修改并添加了卷积层和残差层，整体采用64倍降采样，使其速度大大提升；最后，根据SAR图像目标的特点，改进了损失函数，从而得到了SAR YOLO 960算法。在手工制作的高分辨率SAR图像数据集中的目标检测结果表明，相对于当前主流的检测算法，该算法性能显著提高；检测速度达32.8帧/秒，准确率达95.7%，召回率达94.5%。 |
| 关键词: 损失函数 多尺度检测 YOLOv3网络 残差网络 深度学习 |
| DOI：DOI:10.3969/j.issn.1672-2337.2019.05.018 |
| 分类号:TN957.5 |
| 基金项目:国家自然科学基金青年基金(No.41201468)；湖南省教育厅优秀青年项目(No.16B004) |
|
| Research on SAR Target Detection Method Based on Deep Learning |
|
LIANG Yiqing，WANG Xiaohua，CHEN Lifu
|
|
School of Electrical and Information Engineering，Changsha University of Science and Technology，Changsha 410114，China
|
| Abstract: |
| The sizes of various targets in high resolution synthetic aperture radar (SAR) images are quite different. This makes the characteristics of small targets not obvious，which brings great challenges to target detection. In response to this problem，the SAR YOLO 960 algorithm is proposed. The algorithm first improves the image input size limit and raises the input image to 960×960 pixels. Thus the overall structure of the YOLOv3 (You Only Look Once v3) network is improved. And the convolution layer and the residual layer are modified and added. Using 64 times down sampling，the speed is greatly increased. Finally，according to the characteristics of SAR image target，the loss function is improved. As a result，the SAR YOLO 960 algorithm is obtained. The target detection results in the hand made high resolution SAR image dataset show that the performance of the algorithm is significantly improved compared with the current mainstream detection algorithms；the detection speed is 32.8 frames per second，the accuracy rate is 95.7%，and the recall rate is 94.5%. |
| Key words: loss function multi scale detection YOLOv3 network residual network deep learning |
