用于复杂场景裂缝分割的多层次特征提取算法

A Multi-level Feature Extraction Algorithm for Crack Segmentation in Complex Scenes

针对目前裂缝识别算法大多基于单一自有数据集进行识别,在更换数据集之后效果较差、难以适应复杂场景检测的问题,提出了一种适用于复杂场景的裂缝分割算法,该算法基于卷积神经网络(Convolutional Neural Network,CNN),在HED网络的基础上增加了卷积层,并在各卷积层中增加了批标准化层,应用引导滤波对分割结果细化,取得了更完善精准的分割。验证结果表明,与Ground Truth平均交并比(mean Intersection over Union,mIoU)达到86.9%,且处理速度与其他算法相比具有一定优势。综合对比验证了算法的优越性及可行性,为解决裂缝识别提供了新的方法。

At present,most of the crack identification algorithms are based on a single self-owned data set,and the effect is poor after the data set is replaced,which is difficult to adapt to the problem of complex scene detection.A crack segmentation algorithm for complex scenes is proposed.Based on convolutional neural network,the algorithm adds convolution layer on the basis of HED network,and adds batch standardization layer in each convolution layer.Guided filtering is applied to refine segmentation results and achieve more perfect and accurate segmentation.Through experiments,the average intersection ratio with Ground Truth reaches 86.9%,and the processing speed is better compared with other algorithms.The superiority and feasibility of the proposed algorithm are verified by comprehensive comparison,which provides a new method for crack identification.