高阶特征聚合与卷积调制的SAR图像舰船检测

Ship detection in SAR images based on high-order feature aggregation and convolutional modulation

针对合成孔径雷达(SAR)舰船图像在目标长宽比失衡和小目标聚集等复杂场景中检测困难的问题,该文提出了一种关于海洋舰船检测的目标检测算法,以YOLOv7-Tiny算法为基础采用高阶特征聚合的方法来保留更多小尺度舰船的特征信息,引入卷积调制机制加强颈部网络的特征增强能力,动态非单调聚焦机制边界框损失函数与自适应锚框的有效结合,降低了低质量舰船的检测难度。在遥感舰船高分辨率SAR图像数据集上的实验结果表明,该算法的检测精度和模型参数量分别为90.5%和6.12MB,在SAR舰船检测数据集上进行泛化测试,检测精度能达到97.2%。该文的算法模型提升了小尺度样本的检测能力,相比基准模型具有更高的检测精度和更好的鲁棒性,与其他目标检测算法相比拥有更好的检测效果。

Aiming at the challenges of detecting synthetic aperture radar(SAR)ship images in complex scenes characterized by target aspect ratio imbalance and small target aggregation,this paper proposed a target detection algorithm for marine ship detection.The algorithm was based on the YOLOv7-Tiny algorithm and employed a higher-order feature aggregation method to retain more feature information of small-scale ships.The convolutional modulation mechanism was introduced to enhance the feature enhancement ability of the neck network.The effective combination of the dynamic non-monotonic focusing mechanism bounding box loss function and adaptive anchor frame reduced the difficulty of detecting low-quality ships.Experimental results on the remote sensing ship high-resolution SAR images dataset demonstrated that the algorithm achieved a detection accuracy of 90.5%and had a model parameter count of 6.12 MB.Generalization testing on the SAR ship detection dataset showed a detection accuracy of 97.2%.The algorithm model in the paper improved the detection capability of small-scale samples,exhibiting higher detection accuracy and better robustness compared to the benchmark model.It also exhibited superior detection performance compared to other target detection algorithms.