基于EfficientDet的无预训练SAR图像船舶检测器

针对多尺度、多场景的合成孔径雷达(SAR)图像船舶检测问题，提出了一种基于EfficientDet的无预训练目标检测器。现有的基于卷积神经网络的SAR图像船舶检测器并没有表现出其应有的出色性能。重要原因之一是依赖分类任务的预训练模型，没有有效的方法来解决SAR图像与自然场景图像之间存在的差异性；另一个重要原因是没有充分利用卷积神经网络各层的信息，特征融合能力不够强，难以处理包括海上和近海在内的多场景船舶检测，尤其是无法排除近海复杂背景的干扰。SED就这2个方面改进方法，在公开SAR船舶检测数据集上进行实验，检测精度指标平均准确率(AP)达到94.2%，与经典的深度学习检测器对比，超过最优的RetineNet模型1.3%，在模型大小、算力消耗和检测速度之间达到平衡，验证了所提模型在多场景条件下多尺度SAR图像船舶检测具有优异的性能。 

Ship detector in SAR images based on EfficientDet without pre-training

Aiming at the problem of multi-scale and multi-scene Synthetic Aperture Radar (SAR) ship detection, an object detector without pre-training based on EfficientDet is proposed. The existing SAR image ship detectors based on convolutional neural networks do not show excellent performance that it should have. One of the important reasons is that they depend on the pre-training model of the classification tasks, and there is no effective method to solve the difference between the SAR image and the natural scene image. Another important reason is that the information of each layer of the convolutional network is not fully utilized, the feature fusion ability is not strong enough to deal with the detection of ships in multiple scenes including sea and offshore, and especially the interference of complex offshore background cannot be ruled out. SED improves the method in these two aspects, and conducts experiments on the public SAR ship detection data set. The detection accuracy index AP of SED reaches 94.2%, which, compared with the classic deep learning detector, has exceeded the best RetineNet model by 1.3%, and achieved a balance among model size, computing power consumption and detection speed. This verifies that the model can achieve excellent performance in multi-scale SAR image ship detection in multiple scenes.