基于深度学习的网格优化裂缝检测模型研究

Research on grid optimized crack detection model based on deep learning

为提高混凝土结构裂缝识别的有效性和准确性,基于深度学习技术,建立了一种包含网格优化模块的卷积神经网络(Grid-Deeplab)模型,用于识别裂缝区域并实现像素级语义分割。采用Grid-Deeplab模型,通过对网格化后裂缝图像的具有不同重要性特征的子区域进行建模,使模型具有区别图像有效区域的能力,从而显著提升了裂缝检测模型的检测效率和准确率。采用平均交并比作为评价标准,在数据集上将提出的Grid-Deeplab模型与既有的U-Net、SegNet、FCN和Deeplab等多种神经网络模型进行测试,结果表明,Grid-Deeplab优化模型在测试集上的平均交并比达到0.77,识别准确率优于现有的其他模型。

In order to improve the efficiency and accuracy of crack detection of concrete structure,A convolutional neural network(Grid-Deeplab)including a grid optimization module based on deep learning technology is proposed to identify crack areas and achieve pixel-level semantic segmentation.The Grid-Deeplab model is used to model the sub-regions of the cracked image with different importance features.The model has the ability to distinguish the effective area of the image,thereby significantly improve the efficiency and accuracy of the crack detection model.Using the mean intersection over union(MIoU)as the evaluation standard,the proposed Grid-Deeplab model is tested through the data set with the existing U-Net,SegNet,FCN and Deeplab models.The results show that the MIoU of the Grid-Deeplab optimization model on the test set reaches 0.77,and the recognition accuracy and training efficiency of the model are superior to other existing models.