基于ConSinGan模型的探地雷达病害样本生成方法
Ground Penetrating Radar Disease Image Sample Generation Method Based on ConSinGan Model

探地雷达路基病害智能识别是探地雷达工程探测领域中的一个重要研究分支。针对目前识别任务中缺乏公开数据集，病害样本数量较少和病害类别之间样本不均衡的问题提出基于ConSinGan对抗式生成网络模型的病害样本生成算法MConSinGan。首先，算法利用特征选择的思想设计特征学习模块，并将该模块添加到生成器的头部和尾部，加强不同阶段的生成器之间的联系，提高生成器网络对主要病害特征的学习能力，提取特定区域的有效信息，解决最初模型在生成病害样本的过程中存在的特征不完整的问题；其次，根据每个阶段中图像的放缩比例，以分段函数的形式动态改变每个阶段的迭代次数，平衡因添加特征学习模块而对网络整体增加的训练开销，同时提高生成样本之间的差异性；最后，将改进后的模型应用于实际探地雷达病害图像生成问题中，通过对比实验和消融实验对生成结果进行定量和定性分析。另一方面，将原模型与本文改进模型的训练开销和每阶段训练的平均速度进行比较，从模型训练性能的角度对模型进行评估。实验结果表明，所提算法在探地雷达路基病害图像生成问题中，其生成图像的SIFID值较原模型平均降低了26.5%，MS-SSIM值平均升高了75.7%，在保证训练开销增幅不大的情况下，兼顾生成数据的真实性和生成样本间的多样性。证明了该算法在探地雷达病害样本数据生成问题中能够取得较好的效果。
Intelligent recognition of ground penetrating radar (GPR) subgrade damage is an important research branch in the field of GPR engineering detection. Aiming at the problems of lack of public data set in current recognition tasks, small number of disease samples and unbalanced samples among disease categories, a disease sample generation algorithm MConSinGan based on ConSinGan antagonistic generation network model was proposed. First, the algorithm uses the idea of feature selection to design a feature learning module, and adds the module to the head and tail of the generator to strengthen the connection between generators at different stages, improve the learning ability of the generator network to the main disease features, extract effective information in a specific region, and solve the problem of incomplete features in the process of generating disease samples in the initial model. Secondly, according to the scaling ratio of images in each stage, the number of iterations in each stage is dynamically changed in the form of piecewise function, which balances the increased training cost of the whole network due to the addition of feature learning module and improves the difference between generated samples. Finally, the improved model is applied to the actual ground-penetrating radar disease image generation problem, and the generated results are analyzed quantitatively and qualitatively through comparison experiment and ablation experiment. On the other hand, the training cost and the average training speed of each stage were compared between the original model and the improved model, and the model was evaluated from the perspective of training performance. Experimental results show that the proposed algorithm reduces the SIFID value of the generated image by 26.5% and increases the MS-SSIM value by 75.7% compared with the original model in the problem of ground- penetrating radar subgrade disease image generation. Under the condition that the training cost is not greatly increased, the