卷积神经网络识别侧扫声呐影像的研究

Research on Convolutional Neural Network Model for Image Recognition in Side Scan Sonar

卷积神经网络是当前应用最广泛的图像识别算法,利用大量数据对网络进行训练,即可达到快速、高效地对图像进行识别的目的。由于卷积神经网络结构众多,且同一数据在不同网络表现也不尽相同,为了选择适用于分析沉船声呐数据的网络,使用Python搭建LeNet-5、AlexNet、VGG、GoogLeNet、ResNet系列(ResNet50、ResNet101、ResNet152)、DenseNet系列(DenSeNet121、DenSeNet169、DenSeNet201、DenSeNet264)六种共11个卷积神经网络结构,以侧扫声呐沉船影像为数据集进行实验。结果表明,DenSeNet121是6种网络中最适合用于水下沉船图像识别的网络结构。

Convolutional neural network is currently the most widely used image recognition algorithm.It can achieve the purpose of image recognition quickly and efficiently by training the network with a large amount of data.However,the convolutional neural network has numerous structures,and the same data has different performance in different network.In order to explore the suitable popular network for sunken sonar data,eleven convolutional neural network structure,which are LeNet-5,AlexNet,VGG,GoogleNet,ResNet50,ResNet101,ResNet152,DenSeNet121,DenSeNet169,DenSeNet201 and DenSeNet264,are constructed by Python.Further,the shipwreck images of side-scan sonar are used as the data set for the experiment.The results show that DenSeNet121 is the most suitable network structure for underwater sunken ship image recognition among the those networks.