农村路面多类型病害检测方法研究

Research on the detection method of multi-type diseases on rural pavement

针对实际采集场景下路面影像中病害受背景纹理噪声影响程度大、病害边缘模糊导致分割不准确的问题,该文提出了一种基于Res_UNet和全连接条件随机场的路面病害像素级精准检测方法:(1)对路面影像进行灰度化、中值滤波和自适应直方图均衡化等预处理;(2)根据辽宁省多年份实测路面影像制作大规模、多场景、像素级路面病害数据集,然后融合注意力机制及Dense Crf优化Res_UNet网络结构完成模型训练;并引入损失函数dice loss增强了该方法对细小病害提取的能力;(3)将深度卷积神经网络分割后的路面病害特征图导入全连接条件随机场,对预测的路面病害结果进行轮廓优化,其检测结果为获取路面裂缝宽度,进而评估路面病害等级奠定了基础。该文选用2000张辽宁省农村公路实测路面影像,并以人工判读作为标准,分别从准确率、召回率和精确率3个方面验证本文方法、分水岭算法和Res_UNet模型在实际工作环境下的农村公路路面病害分割性能。结果表明,方法的准确率为91.3%,召回率为87.8%,精确率为87.5%,路面病害轮廓提取更加精细,能够适应于复杂路面条件下病害高鲁棒分割。

Aiming at the problems that the disease in the road image is greatly affected by the background texture noise in the actual collection scene,and the blurred edge of the disease leads to inaccurate segmentation,a road surface based on Res_UNet and Dense Conditional random field(Dense Crf)is proposed.Pixel-level accurate detection method of disease.The method firstly preprocesses the pavement image with grayscale,median filtering and adaptive histogram equalization;secondly,it creates a large-scale,multi-scene,pixel-level pavement disease dataset based on the measured pavement images in Liaoning Province for many years.Integrate the attention mechanism and Dense Crf to optimize the Res_UNet network structure to complete the model training;and introduce the loss function dice loss to enhance the method’s ability to extract small diseases;finally,the road disease feature map segmented by the deep convolutional neural network is imported into the full connection condition The random field is used to optimize the contour of the predicted pavement disease results,and the detection results lay the foundation for obtaining the pavement crack width and then evaluating the pavement disease level.This paper selects 2000 measured pavement images of rural roads in Liaoning Province,and uses manual interpretation as the standard to verify the method,the watershed algorithm and the Res_UNet model in the rural road pavement in the actual working environment from the three aspects of accuracy,recall and precision.The performance of disease segmentation;the results show that the accuracy of the method in this paper is 91.3%,the recall rate is 87.8%,and the precision is 87.5%.The extraction of road disease contours is more refined and can be adapted to highly robust disease segmentation under complex road conditions.