基于生成对抗网络和迁移学习的排水管道缺陷识别

Sewer Defects Recognition Based on Generative Adversarial Networks and Transfer Learning

城市排水设施的服役可靠性成为越来越重要的工程问题,高效率、自动化、大规模的管道缺陷智能检测是城市排水设施建设和管理的迫切需求和重要发展趋势。近年来深度学习技术发展迅速,为排水管道缺陷检测提供了新方法。然而,数据量不足和样本不均衡是深度学习模型普遍存在的问题,影响模型的泛化能力和识别鲁棒性。基于当前先进的生成对抗网络(StyleGAN),提出了一种高质量的排水管道缺陷图像合成方法,以解决训练样本问题。进一步采用卷积神经网络算法,借助迁移学习和预训练模型(SqueezeNet网络)实现管道缺陷识别,提升模型识别效率,并对合成图像进行效果验证。结果表明,StyleGAN能高效合成高质量的缺陷图像,识别模型的平均精度达到90.0%(对树根、错口、残墙坝根和障碍物的精度分别为99.7%、92.3%、87.7%和81.7%)。借助生成对抗网络实现数据增强,为深度学习模型训练提供了一种有前景的方法,具有重要的应用意义。

The service reliability of urban drainage facilities is becoming an increasingly important engineering issue.Highly efficient,automated and large-scale intelligent detection of sewer defects is an urgent need and an important development trend for the construction and management of urban drainage facilities.Deep learning technology has developed rapidly in recent years,providing new methods for sewer defects detection.However,insufficient amount of data and unbalanced samples are common problems of deep learning models,which affects the generalization ability and recognition robustness of the models.Based on the current state-of-the-art generative adversarial network(StyleGAN),a high-quality sewer defects image synthesis method is proposed to solve the training sample problem.Further,a convolutional neural network algorithm is used to implement the sewer defects recognition with the help of transfer learning and pretrained model(SqueezeNet network),and the effect of the synthesized images is verified.Results showed that StyleGAN could efficiently synthesize high-quality sewer defects images with mean average precision of 90.0%for the recognition model(99.7%,92.3%,87.7%and 81.7%for the specific precisions of tree root,disjoint,residential wall and obstacle,respectively).Data augmentation with the help of generative adversarial networks provides a promising approach for deep learning model training with important applications.