seabet bettingTitle:Research on ground penetrating radar seabet betting based on generative adversarial network GAN
seabet betting time:19:30 pm on November 30, 2020
Reporting location:Railway Comprehensive Laboratory Building 308
Reporter:Hou Feifei
seabet betting Summary:
Ground Penetrating Radar, GPR) Profile data collection is the basis of GPR data analysis，Whether it is training data for deep learning feature extraction，It is still used for parameter inversion calculation of underground targets，All require sufficient GPR data。If the data set is too small，Will cause instability in the training network、Inaccurate parameter inversion calculation and other problems。Due to electromagnetic wave signal attenuation，Uneven underground media and other factors，GPR images tend to be distorted、Dislocation、Missing and other phenomena，brings certain difficulties to GPR data interpretation work，Traditional image processing methods are not very applicable。There are almost no public GPR data sets at present，Under actual measurement，Collect and store large、Heterogeneous radar data is quite time-consuming。Given Generative Adversarial Networks, GAN) is a new image generation model based on deep learning，It proposes a new training method based on game theory。Compared to traditional image generation methods，GAN avoids the complex data modeling process，Through powerful deep neural network，Continuously approach the data distribution of real data。This seabet betting describes the application of GAN ideas to GPR data production，Supervised training model based on Auxiliary Classifier GAN (ACGAN) network and gradient optimization idea of Wasserstein GAN-Gradient Penalty (WGAN-GP)，And adopt three different ResNet block structures (upsampling，Downsampling，Convolution) improves network performance，Complete two parallel tasks of image classification and image confidence at the same time。The proposed GAN network can quickly generate GPR profiles by simulating the distribution of the model，Solved the complexity of GPR data measurement to a certain extent，Simulation time-consuming and other issues，It also provides a new idea for future radar data collection。
