基于CNN海上钻井平台检测模型的构建及训练算法分析

卷积神经网络(CNN)是深度学习(DL)中最具代表性的一种网络结构。合成孔径雷达(SAR)图像具有位置结构关系,CNN模型可以利用图像的位置结构关系,能够更好地提取图像特征,因此更适合采用CNN模型检测海洋目标。本文首先基于CNN框架构建了海上钻井平台检测的DL模型Ocean TDAx,并对模型进行训练和测试。试验结果表明,Ocean TDA9模型精度最高。然后针对Ocean TDA9模型,采用Adam、RMSprop、Stochastic gradient descent (SGD)、Adagrad、Momentum等7种模型训练算法进行试验,比较不同算法的训练损失和精度与训练批次的相关性。最后基于渤海海域的极化SAR数据,对提出的Ocean TDA9模型、已有的CNN模型及VGG模型进行海上钻井平台检测对比。结果表明,构建的Ocean TDA9模型在钻井平台检测中整体性能优良。

Detection model construction based on CNN for offshore drilling platform and training algorithm analysis

Convolutional neural networks (CNN) is the most representative network structure of deep learning (DL). Synthetic aperture radar (SAR) image itself has the position structure relationship, the characteristics of the CNN model determine that it can use the position structure relationship of the image to extract the features of the image better,so it is more suitable to use the CNN model for marine target detection.The paper based on CNN framework constructs the DL model Ocean TDAx of offshore drilling platform detection, trains and tests the OceanTDAx model through the improved WinR-Adagrad gradient training algorithm, experimental results show that the Oceant TDA9 model is the highest accuracy. For the Ocean TDA9 model, seven model training algorithms, such as adam, RMSprop, Stochastic gradient descent (SGD), Adagrad and Momentum, is used to conduct experiments,and the training loss and accuracy with relevance of training batches of different algorithmsis is compared. Based on the polarized SAR data of the Bohai sea, the proposed Ocean TDA9 model and the existing CNN model and Visual geometry group (VGG) model are used to compare the detection experiments of offshore drilling platforms.The results show that the constructed Ocean TDA9 model is excellent overall performance in drilling platform testing.