随着我国公路网的不断扩展,道路病害检测已经成为道路养护与行车安全保障必不可少的组成部分,基于深度学习的道路病害检测已经成为该领域的研究热点.针对多种病害并发的复杂场景下道路病害识别精度不高、泛化能力不足的问题,提出了一种复杂场景下的道路病害检测模型RGT-YOLOv7(Receptive Ghost Triplet-YOLOv7).在主干网络部分引入三重注意力机制,提高病害特征在不同通道与空间的相关性,解决了特征提取效率不高的问题;将原有的全连接空间金字塔卷积模块替换为快速全连接空间金字塔卷积模块,并加入幻影卷积模块,提高冗余特征的使用率,将原有的冗余特征与新提取到的特征融合,得到包含不同尺度的特征信息;为了扩大模型感受野,在特征增强部分加入改进的感受野模块,利用不同尺寸的空洞卷积从不同方向对特征图进行提取,加强对横向和纵向特征的提取.实验结果表明,与YOLOv7相比,识别的平均正确率(mean average precision,mAP)和平衡F分数分别提升了6.9、3.9个百分点,尤其是对纵向裂缝危害识别的平均正确率提高了22.3个百分点,与Faster R-CNN、YOLOv5等模型相比也有良好的性能提升,表明RGT-YOLOv7是一种有效的复杂场景下的道路病害检测模型.
Road Disease Detection RGT-YOLOv7 Model under Multiple Diseases Complicated Scenarios
With the continuous expansion of China's road network,road disease detection has become an indispensable part of road maintenance and traffic safety,and road disease detection based on deep learning has become a research hotspot in this field.Aiming at the problems of low accuracy and generalization ability of road disease identification in complex scenes with multiple diseases,a road disease detection model called Receptive Ghost Triplet-YOLOv7(RGT-YOLOv7)in complex scenes is proposed in this paper.A triplet attention mechanism is introduced in the backbone network to improve the correlation of disease features in different channels and spaces,and to solve the problem of low feature extraction efficiency.The original SPP module is replaced by the SPPF module,the Ghost module is added to improve the utilization rate of redundant features,and the original redundant features and the newly extracted features are fused to get more diverse and rich feature information with different scales.In order to improve the model perception field,RFBs module is added in the feature enhancement part,and the feature map is extracted from different directions by using cavity convolution with different sizes to enhance the extraction of horizontal and vertical features.Experimental results show that the average accuracy and balanced F score are improved by 6.9 percentage points and 3.9 percentage points,respectively,compared with YOLOv7,especially the longitudinal fracture identification is improved by 22.3 percentage points,and it also has good performance improvement compared with Faster R-CNN,YOLOv5,and recently proposed algorithm models.It is an effective road disease detection algorithm for proposed RGT-YOLOv7 under complex scenes.
object detectionroad distress detectiondeep learningYOLOv7receptive field block(RFB)
国家科技期刊平台
NSTL
万方数据
