基于图像的道路裂缝分割及量化方法研究

Research on image-based segmentation and quantification of road cracks

针对当前道路裂缝检测与量化领域中高成本与高精度之间的矛盾,本文提出了一种低成本高精度的道路裂缝自动分割量化系统。首先,该系统采用添加跳跃级往返多尺度融合模块与注意力门机制的卷积神经网络进行分割预测,并命名为SW-Net。随后,通过结合MCO、DFS以及不同方向上像素统计曲线的变化趋势对裂缝进行分类。最后,为了克服裂缝量化的不连续性和传统形态学骨架量化算法的局限性,本文通过结合A*算法并对其进行扩展,以计算裂缝的最短长度和最大宽度。实验对比结果表明,该系统在Crack500数据集上取得了所有对比模型中最佳的准确率(93.68%)和F1分数(0.8965)。改进后分类算法的平均分类精度达到99.29%,分类速度为109张/s。其量化最短长度和最大宽度的相对误差分别为12.34%和15.85%,较传统骨架方法的平均量化误差降低了5.16%。这些结果表明,该系统在裂缝的分割、分类和量化方面取得了显著进展。

Aiming at the contradiction between high cost and high precision in the field of road crack detection and quantification,this paper proposes a low cost,high precision automatic segmentation and quantification system for road cracks.Firstly,the convolutional neural network with jump-stage round-trip multi-scale fusion module and attention gate mechanism is used for segmentation prediction,which is named SW-Net.Then,the cracks are classified by combining MCO,DFS and the trend of pixel statistical curves in different directions.Finally,in order to overcome the discontinuity of crack quantization and the limitation of traditional morphological skeleton quantization algorithm,this paper combined the A algorithm and extended it to calculate the shortest length and maximum width of cracks.Experimental comparison results show that the system achieves the best accuracy(93.68%)and F1 score(0.8965)among all comparison models on the Crack500 dataset.The average classification accuracy of the improved classification algorithm is 99.29%,and the classification speed is 109 pieces/s.The relative errors of the shortest length and maximum width are 12.34%and 15.85%respectively,which is 5.16%lower than the average error of the traditional skeleton method.These results show that the system has made remarkable progress in the segmentation,classification and quantification of cracks.Aiming at the contradiction between high cost and high precision in the field of road crack detection and quantification,this paper proposes a low cost and high precision automatic segmentation and quantification system for road cracks.Firstly,the convolutional neural network with jump-stage round-trip multi-scale fusion module and attention gate mechanism is used for segmentation prediction and named SW-Net.Then,the cracks are classified by combining MCO,DFS and the trend of pixel statistical curves in different directions.Finally,in order to overcome the discontinuity of crack quantization and the limitation of traditional morphological skeleton quantization algorithm,this paper combined the A algorithm and extended it to calculate the shortest length and maximum width of cracks.Experimental comparison results show that the system achieves the best accuracy(93.68%)and F1 score(0.8965)among all comparison models on the Crack500 dataset.The average classification accuracy of the improved classification algorithm is 99.29%,and the classification speed is 109 pieces/s.The relative errors of the shortest length and maximum width are 12.34%and 15.85%respectively,which is 5.16%lower than the average error of the traditional skeleton method.These results show that the system has made remarkable progress in the segmentation,classification and quantification of cracks.