基于自监督预训练的小样本道路病害检测

Few-shot Road Disease Detection Based on Self-supervised Pre-training

在交通道路病害检测领域经常遇到小样本问题,导致深度学习模型训练困难。为了解决在道路小样本数据上监督信息不足的问题,提出了一种基于自监督预训练的小样本分类框架。首先能充分利用所有数据,包括无标注的样本,根据其数据变换的方式形成伪标签,利用伪标签对特征提取器进行预训练;再将特征提取器参数迁移到有监督模型中,利用带真实标注的小样本数据进行模型的微调,最终将用于道路病害分类。另外,该框架可以使用不同的特征提取器和小样本度量方法。实验使用真实的小样本道路数据对该自监督预训练框架进行测试和分析,证明能够有效地提升病害检测准确率。

In the field of traffic road disease detection,problems of learning from few samples are often encountered,leading to the insufficient training of deep learning models.In order to solve the lack of supervisory signal with few-shot road data,a few-shot classification framework based on the self-supervised pre-trained model is proposed.Firstly,this framework can make full use of the all available data,including unlabeled samples,by forming pseudo labels according to the way of data augmentation.Then,the feature extractor can be pre-trained with pseudo labels.Secondly,the pre-trained parameters are transfered to the super-vised model,and the few-shot data with ground-truth labels is applied to fine-tune the whole model.Finally,the model will be used for road disease classification.In addition,the framework can be compatible with various feature extractors and few-shot met-ric measurements.The experiments use the real few-shot road data to test and analyze the self-supervised pre-training framework,which proves that this framework can effectively improve the accuracy of disease detection.