改进YOLOv3的SAR图像舰船目标检测

针对目标检测算法直接应用于SAR图像舰船检测数据集时数据训练不充分、鲁棒性差等问题，本文提出了一种改进YOLOv3的SAR图像舰船目标检测方法，从改进网络训练策略的角度出发，提升算法对不同舰船目标的适应性，优化算法的检测性能。改进主要包括两个方面:一方面本文在YOLOv3的基础上引入了ATSS（Adaptive Training Sample Selection）正负样本的分配方法，提高YOLOv3中正负样本选择的质量，优化网络训练。另一方面本文设计了基于特征层的锚框超参数优化方法，使锚框更加贴合各检测层数据集样本分布，从而使训练模型更好的收敛。本文分别在SSDD、SAR-Ship-Dataset数据集上进行了实验，验证了其有效性。 

Ship target detection in SAR image based on improved YOLOv3

Aiming at the problems of insufficient data training and poor robustness when target detection algorithm was directly applied to SAR image ship detection, this paper proposed a ship detection method based on improved yolov3. From the perspective of improved network training strategy, it improved the adaptability of the algorithm to different ship targets and optimized the detection performance of the algorithm.The improvement of this method included two aspects: on the one hand, this paper introduced the distribution method of positive and negative samples of ATSS based on yolov3. This method improved the imbalance between positive and negative samples and improved the quality of training sample selection.On the other hand, this paper designed the anchor boxes’ super parameter optimization method based on the feature layer, which made the anchor boxes more fit the samples distribution of each detection layer data set, so that the training model converged better.The experimental results were carried out on SSDD and SAR-Ship-Dataset respectively. The improved algorithm has improved the detection performance in inshore and offshore scenes, and verified its effectiveness.