为解决合成孔径雷达舰船检测在复杂背景、小舰船目标和目标舰船体积相差较大的情况下存在虚警、漏警和置信度偏低的问题,提出了一种基于改进可切换空洞卷积的合成孔径雷达舰船检测算法。通过将ELAN 层中的卷积改进为可切换空洞卷积和添加通道注意力机制的方法来扩大卷积层的感受野,高效地聚合网络中不同层的特征信息;在颈部特征融合处加入快速加权特征融合AIFI 模块,提高效率并减少模型的冗余计算量;在损失函数处通过构造梯度增益的计算方法来引入聚焦机制。该算法提高了模型检测小目标的能力,并解决了在复杂背景下虚警、漏警的问题;通过使用SSDD 数据集对改进后的模型进行验证,相较于改进前的基准YOLO-7 模型,改进后的mAP 值达到96. 59??相较基准模型提升了9. 33??,同时准确率和召回率分别提升 3. 81??和16. 36??。实验结果表明,改进后的算法有效提升舰船目标的检测精度,显著改善小目标检测中存在的虚警和漏警问题。

In order to solve the problems of false alarms, missed alarms and low confidence of SAR ship detection in complex backgrounds, small ship targets and large difference in the size of target ships, a SAR ship detection algorithm based on the improvement of the switchable null convolution is proposed. By improving the convolution in the ELAN layer to switchable null convolution and adding the channel attention mechanism to expand the sensory field of the convolution layer, the feature information of different layers in the network is efficiently aggregated; a fast weighted feature fusion AIFI module is added at the neck feature fusion to improve the efficiency and reduce the redundant computation of the model; and a focusing mechanism is introduced by the construction of a gradient gain computation method at the loss function. The algorithm improves the model's ability to detect small targets. This algorithm improves the ability of the model to detect small targets and solves the problem of false and missed alarms in complex backgrounds. The improved switchable null convolutional model is validated by using the SSDD dataset, and compared with the benchmark YOLO-7 model before the improvement, the improved mAP value reaches 96. 59?? compared with the benchmark model, which is an increase of 9. 33??, while the accuracy and recall are increased by 3. 81% and 16. 36%, respectively. The experimental results show that the improved algorithm effectively improves the detection accuracy of ship targets and significantly improves the false alarm and missed alarm problems in small target detection.

TN957. 52

国家自然科学基金资助项目(61201418, 62171293); 深圳市基础研究专项资助项目(JCYJ20230808105359045)