Abstract: Cracks are one of the important diseases on the pavement surface. Traditional crack detection relies on manual visual inspection, which is time consuming and labor intensive. Although traditional image processing techniques can make crack detection and segmentation more automated to some extent. However, image processing techniques are susceptible to some noise caused by illumination, blur, and the like. In order to complete the segmentation and detection of pavement cracks in complex environments, a segmentation method based on improved Fully Convolutional Networks（FCN）is proposed. According to the established data set, the traditional FCN and the optimized FCN are trained. The test results show that the mean Intersection over Union（mean_IoU） is improved, so the proposed method can segment the cracks accurately.

Key words: crack detection, image processing, Fully Convolutional Networks（FCN）, mean intersection over union

摘要： 裂缝是路面表面的重要病害之一。传统的裂缝检测依赖于人工的视觉检查，在实际操作中费时费力。虽然传统的图像处理技术在一定程度上可使裂缝检测与分割更加自动化，但是图像处理技术易受到由光照、模糊等引起的一些噪声的影响。为了完成复杂环境下对路面裂缝的分割及检测，提出了一种基于改进的全卷积神经网络（Fully Convolutional Networks，FCN）的分割方法，根据建立的数据集训练传统FCN和优化后的FCN，测试结果表明其平均交并比（mean_IoU）得到了一定的提高，故该方法能够较准确地分割出路面裂缝。

关键词: 裂缝检测, 图像处理, 全卷积网络（FCN）, 平均交并比