GAN数据增强下路面裂缝语义分割算法

Semantic segmentation algorithm of pavement cracks based on GAN data augmentation

为解决现实情况下路面裂缝图像采集数量无法满足深度学习样本数量要求的问题,基于生成式对抗网络的数据扩增方法,提出一种以改进型U-Net网络模型为基础的路面裂缝语义分割算法。首先,基于传统图像处理方法将采集到的样本数据进行初次扩充,根据生成式对抗网络原理,实现样本数据再次扩充;其次,通过增加网络层数、添加归一化层、添加Dropout层提出一种基于改进型U-Net网络模型的路面裂缝语义分割算法;最后,利用基于改进型UNet网络的路面裂缝语义分割模型提取扩增数据图像中的裂缝,在同等条件下与传统U-Net网络模型检测算法、现有主流分割算法FCN作实验对比研究。结果表明:该改进型算法的分割精度均优于其他两种算法,能较为准确地分割出路面裂缝,在背景像素较为复杂的情况下能较好地避免误检的情况,其平均像素精度、平均交并比分别达到了92.43%、83.43%,在实际场景应用中具有较好的检测效果及较强的泛化性能。

In view of the problem that the number of pavement crack images cannot meet basic needs for deep learning,according to using generative adversary network to expand the data-set,a pavement segmentation algorithm based on the U-Net network is proposed.Firstly,the data-set was initially expanded by traditional image generation,according to the principle of generative adversary network,a algorithm of pavement crack segmentation based on semantic segmentation was proposed,which was used to expand the data-set again.Secondly,based on the U-Net,an algorithm of pavement crack segmentation based on semantic segmentation was proposed,which increased the number of network layers and added Batch Normalization and dropout layer.Finally,the semantic segmentation model of pavement cracks was used to extract cracks in the expanded data image,and compared with the traditional detection algorithm and the existing mainstream segmentation algorithm FCN.The results show that the segmentation accuracy of the algorithm is better than other two algorithms,which more precisely segments pavement crack images and avoids error detection when background pixel is complicated.The mean pixel accuracy and mean intersection over union of the algorithm are 92.43%and 83.43%,respectively.In the practical scene application,it has better detection effect and stronger generalization performance.