复杂背景下小样本声呐图像目标检测

Few-shot sonar images target detection in complex backgrounds

水下目标检测具有重要意义,在军事和民用领域都发挥着重要作用。实际场景中可以获得的声呐图像非常有限,且声呐图像的信噪比较低,无法得到较好的检测结果。因此,本文引入小样本学习,基于Faster RCNN两阶段目标检测算法,选择不同的策略对模型进行优化,得到了较好的检测结果并验证了小样本目标检测在声呐图像领域的可行性。根据混响对声呐图像的影响进行仿真实验,得到不同混响背景下的声呐图像,对比分析了不同数据集下训练模型的检测性能。实验结果表明,在训练样本中增加混响信号可以提高低信噪比条件下的目标检测精度。

Underwater target detection is a research problem of great significance and plays an important role in both military and civilian fields.The sonar images available in real scenes are very limited and the low signal-to-noise ratio of the sonar images does not allow for satisfactory detection results.Therefore,this paper introduces few-shot learning,based on the Faster R-CNN two-stage target detection algorithm,and chooses different strategies to optimize the model,obtaining better detection results and verifying the feasibility of few-shot target detection in the field of sonar images.Then,simulation experiments are conducted to obtain sonar images under different reverberation backgrounds according to the effect of reverberation on sonar images,and the detection performance of the training model under different datasets is compared and analyzed.Experimental results show that adding reverberation signals to the training samples can improve the target detection accuracy under low signal-to-noise ratio conditions.