为同时获取路面裂缝的位置信息与分布路径,以及形状延展和密度信息等,对目标检测算法与图像分割算法的融合进行了研究,在分析了目标检测算法与图像分割算法的网络结构与特征融合方式后提出了一种基于YOLO V5与PSPnet的PSP-YOLO裂缝检测分割算法。同时针对裂缝图像采集困难、样本不足的问题提出一种基于GAN网络的数据扩增网络,生成以假乱真的裂缝图像对裂缝样本进行扩增。试验结果表明,PSP-YOLO检测分割算法能够同时获取裂缝的位置与形状延展等信息,在该数据集下,其路面裂缝检测的平均精度为93.18%,分割模块的平均交并比为74.68%。在相同的试验条件下,所提方法比原YOLO V5的平均精度提高了2.69%,分割模块的平均交并比比原PSPnet的提高了1.54%。

In order to obtain the location information, distribution path, shape extension and density information of pavement cracks at the same time, the fusion of target detection algorithm and image segmentation algorithm is studied. After analyzing the network structure and feature fusion mode of target detection algorithm and image segmentation algorithm, a PSP-YOLO crack detection and segmentation algorithm based on YOLO V5 and PSPnet is proposed. At the same time, a data augmentation network based on GAN network is proposed to generate false fracture images to augment fracture samples. The experimental results show that the PSP-YOLO detection and segmentation algorithm can obtain the information of crack location and shape extension at the same time. The average accuracy of pavement crack detection under this data set is 93. 18%, and the average intersection over union of segmentation module is 74. 68%. Under the same experimental conditions, the average accuracy of the segmentation module is 2. 69% higher than that of the original YOLO V5, and the average intersection over union of the segmentation module is 1. 54% higher than that of the original PSP-net.