基于改进SSD的合成孔径声呐图像水下多尺度目标轻量化检测模型

针对轻量化目标检测模型SSD-MV2对合成孔径声呐(SAS)图像水下多尺度目标检测精度低的问题，该文提出一种新的卷积核模块-可扩张可选择模块(ESK)，ESK具有通道可扩张、通道可选择和模型参数少的优点。与此同时，利用ESK模块重新设计了SSD的基础网络和附加特征提取网络，记作SSD-MV2ESK，并为其选择了合理的扩张系数和多尺度系数。在合成孔径声呐图像水下多尺度目标检测数据集SST-DET上，SSD-MV2ESK在模型参数基本相等的条件下，检测精度比SSD-MV2提升4.71%。实验结果表明，SSD-MV2ESK适用于合成孔径声呐图像水下多尺度目标检测任务。

Synthetic Aperture Sonar Underwater Multi-scale Target Efficient Detection Model Based on Improved Single Shot Detector

In view of the problem that the efficient detection model SSD-MV2 (Single Shot Detector MobileNet V2) has low detection accuracy to underwater multi-scale targets in Synthetic Aperture Sonar (SAS) images, a novel feature extraction module Extended Selective Kernel (ESK) is proposed in this paper. ESK has the advantages of channel scalability, channel selection and few model parameters. At the same time, the basic network and additional feature extraction network of SSD are redesigned by using ESK module, which is named SSD-MV2ESK, and a set of reasonable expansion coefficient and multi-scale coefficient are selected for SSD-MV2ESK. On SST-DET, the mAP of SSD-MV2ESK is 4.71% higher than that of SSD-MV2 when the model parameters are basically the same. The experimental results show that SSD-MV2ESK is suitable for SAR underwater multi-scale target detection task in embedded platform.