基于改进型YOLOv3的SAR图像舰船目标检测

传统合成孔径雷达(synthetic aperture radar, SAR)图像目标检测的方法依赖于人工设计特征且易受复杂背景干扰, 泛化能力较差。深度学习的方法可以自动提取特征且具有良好的抗干扰特性, 对于未来雷达智能感知具有重要意义。不同于其他只能对固定区域进行检测的常规卷积神经网络, 本文提出一种改进型YOLOv3的SAR图像舰船目标检测方法, 该方法基于舰船尺寸与形状自适应采样的可变形卷积、ResNet50变体特征提取器和ShuffleNetv2轻量化思想等设计YOLOv3模型。通过SSDD数据集验证, 在检测效果方面, 相较于原YOLOv3模型, 平均精度从93.21%提高至96.94%, 检测概率从95.51%提高至97.75%;在模型大小方面, 轻量化设计模型仅为原YOLOv3模型的八分之一, 可实现嵌入式的使用。

Ship detection in SAR image based on improved YOLOv3

Traditional synthetic aperture radar(SAR)image target detection methods rely on manual design features and are vulnerable to complex background interference,and their generalization ability is poor.The deep learning method can automatically extract features and has good anti-jamming characteristics,which is of great significance for future radar intelligent perception.Different from other conventional neural networks that can only detect fixed areas,an improved YOLOv3 SAR image ship detection method is proposed in this paper.The new YOLOv3 model is designed based on deforming convolution of adaptive sampling of the ship size and the shape,the ResNet50 variant feature extractor and the ShuffleNetv2 lightweight idea.Through SSDD dataset verification,compared with the original YOLOv3 model,the average accuracy increases from 93.21%to 96.94%,and the detection probability increases from 95.51%to 97.75%.In terms of the model size,the lightweight design model is only one-eighth of the original YOLOv3 model,which can be embedded for use.