SCTD1.0:声呐常见目标检测数据集

SCTD 1.0:Sonar Common Target Detection Dataset

近年来,卷积神经网络(CNN)在大规模自然图像数据集(如ImageNet,COCO)中获得了广泛应用,但在声呐图像检测识别领域的应用研究较缺乏,其存在声呐图像目标检测和分类数据集缺乏且水下目标样本往往面临样本稀少、不平衡等问题。针对这一问题,在进行广泛收集声呐图像的基础上,构建了一个完全公开的、可以用于开展声呐图像检测和分类研究的声呐常见目标检测数据集SCTD1.0,该数据集目前已包含水下沉船、失事飞机残骸、遇难者3类典型目标,共计596个样本。在SCTD1.0的基础上,文中采用迁移学习的方式测试了检测和分类的基准,具体来说:针对检测任务,使用特征金字塔网络对多尺度特征进行组合利用,比较了YOLOv3,Faster R-CNN,Cascade R-CNN这3种检测框架在本数据集上的性能表现;针对分类任务,对比了VGGNet,ResNet50,DenseNet 3种网络的分类性能,分类准确率达到了90%左右。

In recent years,convolutional neural networks(CNN)have been widely used in large-scale natural image datasets(such as ImageNet,COCO).However,there is a lack of applied research in the field of sonar image detection and recognition,which suffers from a lack of sonar image target detection and classification datasets and often faces sparse and unbalanced samples of underwater targets.In response to this problem,based on the extensive collection of sonar images,this paper constructs a completely open sonar common target detection dataset SCTD1.0 that can be used for sonar image detection and classification research.The dataset currently contains three types of typical targets:underwater shipwreck,wreckage of crashed aircraft,and victims,with a total of 596 samples.On the basis of SCTD1.0,this paper uses transfer learning to test the benchmarks of detection and classification.Specifically,for the detection task,the feature pyramid network is used to combine and utilize multi-scale features,and the performance of the three detection frameworks YOLOv3,Faster R-CNN,and Cascade R-CNN on this dataset is compared.For classification tasks,the classification performance of the three networks of VGGNet,ResNet50,and DenseNet is compared,and the classification accuracy rate reaches about 90%.