基于混合损失ResNet34-UNet的路面裂缝分割方法

Research on pavement cracks segmentation method based on mixed loss ResNet34-UNET

路面裂缝图像由于其形状细长、弯曲复杂等特点,在模型训练中存在裂缝样本不平衡问题,为此提出了一种基于混合损失函数的ResNet34-UNet路面裂缝分割方法。该方法借助于U-Net结构,以ResNet-34作为主干提取网络,根据数据集中裂缝像素所占比例对BCEFocal Loss和Tversky Loss进行权重调整,并使用调整后的BCEFocal Loss和Tversky Loss组成混合损失函数,平衡了裂缝样本输入和输出不平衡问题。对比实验表明文中的网络模型的F1分数(0.7018)、MIoU(0.8306)均为最高,说明该分割算法能有效地对路面裂缝进行准确分割。

Due to the characteristics of long and thin shape and complex bending, the pavement crack image has the problem of uneven crack samples in model training. Therefore, a Resnet34-UNET pavement crack segmentation method based on mixed loss function is proposed in this paper. With the help of U-NET structure, ResNET-34 was used as the backbone to extract the network, and the weights of BCEFocal Loss and Tversky Loss were adjusted according to the proportion of crack pixels in the data set. The adjusted BCEFocal Loss and Tversky Loss were used to form a mixed Loss function to balance the imbalance between input and output of fracture samples. Comparative experiments show that the F1 score(0.7018) and MIoU(0.8306) of the network model in this paper are both the highest, indicating that the segmentation algorithm can effectively and accurately segment pavement cracks.