基于卷积神经网络的路面裂缝分割设计与研究

Design and research on pavement crack segmentation based on convolutional neural network

  • 摘要: 裂缝是路面病害最主要的类型,准确的裂缝分割是国家进行公路预防养护管理的重要决策依据。针对背景复杂下现有模型路面裂缝分割准确度有待提高的问题,提出一种基于卷积神经网络的端到端裂缝分割模型,使用分层结构的ConvNeXt编码器提取多尺度特征,特征的最高层使用金字塔池化模块进一步获取全局先验特征,通过具有横向连接和自上而下的金字塔结构进行特征融合。针对裂缝和背景不平衡问题,使用平衡交叉熵损失函数提高模型的检测性能。此外,构建了一个包含2876张裂缝图片的数据集UCrack,覆盖多种裂缝类型和广泛的背景范围,以提供丰富的特征供模型学习。实验表明,在UCrack测试数据集上模型的召回率和F1得分比其他表现最佳的模型提高了2.68%和6.89%;在CrackDataset数据集上的测试取得了85.68%的召回率和80.11%的F1得分,说明模型具有较好的泛化性能,可应对背景复杂的路面裂缝分割。

     

    Abstract: Cracks are the most important type of pavement diseases, and the accurate crack segmentation is an important decision basis for national preventive maintenance management of roads. To address the problem of crack segmentation accuracy of existing models for pavement under complex background, an end-to-end crack segmentation model based on convolutional neural network was proposed, which used a layered structure of ConvNeXt encoder to extract multi-scale features. A pyramid pooling module was used to further obtain the global priori features by the top layer of features, and the feature fusion was performed through a pyramid structure with lateral connections and top-down. A weighted cross-entropy loss function was employed to enhance the detection performance of model for the crack and background imbalance problem. In addition, a crack dataset UCrack with 2 876 cracks covering multiple crack types and a wide range of backgrounds was created to provide rich features for model learning. Experiments show that, compared with other best-performing models, the model recall and F1 score on the UCrack test dataset are improved by 2.68% and 6.89%, respectively. The test on the CrackDataset dataset achieves recall of 85.68% and F1 score of 80.11%, which implies that the model has better generalization capability and can cope with pavement crack segmentation with complicated scenarios.

     

/

返回文章
返回