基于异常数据生成和自监督学习的路面裂缝检测

Road Crack Detection Based on Abnormal Data Generation and Self-supervised Learning

为了确保道路安全,有效识别路面裂缝是至关重要的.但是,由于道路状况的复杂多变,大部分现有研究依然采用有监督学习的方法来检测和定位裂缝.这种方法通常受限于标注了裂缝的数据集的可用性,而人工标注路面图像不仅耗时而且成本高昂.针对上述问题,本研究提出了一个简单而有效的基于自监督学习的两阶段分类神经网络,仅使用正常的道路表面数据进行图像异常检测.首先通过使用CutPaste模块处理正常的道路表面数据,生成异常数据.然后,利用经过迁移学习增强的Resnet18网络提取特征,并进行归一化处理以产生分类结果.训练后的网络通过结构剪枝保持了其特征提取能力,并将参数数量减少了74.8%.同时与剪枝前的网络相比,其识别准确度仅下降了1.7%.训练后的Tiny-ResNet18模型参数仅有2.97M,在Crack Forest Dataset和Deep Crack dataset上分别达到了95.30%和98.04%的AUC,实现了高识别精度.

To ensure road safety,it is crucial to effectively identify road cracks.However,due to the complex and variable conditions of roads,most existing research still adopts supervised learning methods for crack detec-tion and localization.These methods are typically limited by the availability of datasets labeled with cracks,and manually annotating road images is not only time-consuming but also costly.In response to the above issues,this paper proposes a simple yet effective two-stage classification neural network based on self-supervised learning that uses only normal road surface data for image anomaly detection.Initially,abnormal data is generated by pro-cessing normal road surface data using the CutPaste module.Subsequently,features are extracted using the Resnet18 network enhanced through transfer learning,followed by normalization to produce classification re-sults.The trained network maintains its feature extraction capabilities through structural pruning,reducing the number of parameters by 74.8%.At the same time,compared to the network before pruning,the recognition accu-racy has only decreased by 1.7%.The trained Tiny-ResNet18 model has 2.97M parameters and achieved AUCs of 95.30%and 98.04%on the Crack Forest Dataset and Deep Crack dataset,respectively,demonstrating high rec-ognition precision.