基于卷积神经网络的钢箱梁裂缝检测算法研究

Research on crack detection algorithm of steel box girder based on convolutional neural network

地震作用和车辆动载的日常冲击常常会引起钢箱梁的表面裂缝损伤。在日常检修和震后开展安全检查时,由于裂缝在整张图像中的占比较低且受到笔迹、锈迹和焊缝等因素的严重干扰,很难从现场采集到的图像中高效并精确地检测出裂缝。为此,本文基于卷积神经网络提出了一种结合裂缝定位和裂缝分割的级联裂缝检测模型。首先,采用分类模型在原始图像中定位出裂缝所在位置,然后采用U-Net模型实现对裂缝的像素级检测。结果表明:本文方法可以实现对大部分裂缝的精确检测,F 1分数达到0.67,高于仅采用分割模型的0.55;此外,相比仅采用分割模型的检测方法,本文方法的检测效率提高了近70%,可达到9.25 s每张。

Earthquake and daily impact of vehicle dynamic load often cause surface crack damage on steel box girders.Due to the low proportion of the crack area in the whole image and the serious interference of factors such as handwriting,rust,welds,etc.,it is hard to detect cracks from the images collected from the scenes is a very difficult task during daily maintenance and post-earthquake safety inspection.In this paper,a cascade crack detection method which combines crack location and crack segmentation model based on convolutional neural networks was proposed.First,a classification model was used to locate the region of the crack in the original image,and then the U-Net model was applied to accurately extract the crack in pixel-level.The detection results showed that the proposed method can achieve accurate extraction of most cracks.The F1 score of the proposed method reached 0.67,which was higher than that of only using segmentation model.In addition,comparing to the method only using segmentation model,the detection efficiency of the proposed method improved nearly 70%,reaching 9.25 seconds per image.