基于两种U型网络的钢铁图像缺陷检测方法对比

Comparison of defect detection methods of steel image based on two U network models

针对钢铁图像缺陷检测问题,使用U型的深度学习神经网络模型U-Net和ResUNet对有缺陷的钢铁图像进行图像分割。通过python对钢铁图像进行数据统计,分析每类缺陷占比,以及缺陷数量与缺陷面积的关系。通过数据生成器依次将数据输入U-Net和Res-UNet模型,对比在相同损失函数下的Tversky系数和损失率。实验结果表明,Res-UNet模型的Tversky系数优于U-Net模型,且Res-UNet模型损失率低于U-Net模型,Res-UNet模型可以更准确地预测钢铁缺陷位置和缺陷类别。

Aiming at the problem of steel image defect detection,the deep learning neural network models U-Net and Res-UNet were used to segment the steel image.The proportion of each kind of defects,the relationship between the number of defects and the area of defects were analyzed through the data statistics of steel images by python.Data generators were used to input data into U-Net and ResUNet models in turn,and the Tversky coefficient and loss rate under the same loss function were compared.The experimental results show that the Tversky coefficient of Res-UNet model is better than that of U-Net model,and the loss rate of Res-UNet model is lower than that of U-Net model.Res-UNet model can predict the location and type of steel defects more accurately.