改进DETR的无人机航拍图像沥青路面破损检测算法

Improved DETR Algorithm for Asphalt Pavement Damage Detection in UAV Aerial Images

针对航拍沥青路面图像数据不足、检测精度低、存在漏检的问题,研究提出一种改进的DETR(Detection Transformer)端到端沥青路面破损检测模型。该模型采用ResNet50提取特征,引入SiLU激活函数提高特征提取能力,并采用多尺度融合特征图保留更多上下文语义信息;在Transformer的Encoder中使用多尺度可变形自注意力机制,加快模型收敛速度;采用CIoU损失函数提高了裂缝检测的准确性。实验结果表明:改进模型的平均精度达83.7%,比DETR模型在精确率上提高7.4%,召回率上提升了10.9%。提出的改进模型可对沥青路面破损进行有效检测,可为航拍图像的沥青路面破损检测提供参考。

Aiming at the problems of insufficient data,low Detection accuracy and missed detection of aerial images of asphalt pavement,an improved DETR(Detection Transformer)end-to-end asphalt pavement damage detection model is proposed.Firstly,the model uses ResNet50 to extract features,introduces the SiLU activation function to improve feature extraction ability,and uses a multi-scale fusion feature map to retain more context semantic information.Sec⁃ondly,the multi-scale deformable self-attention mechanism is used in the Transformer Encoder to accelerate the con⁃vergence speed of the model.Finally,the CIoU loss function is used to improve the accuracy of crack detection.The ex⁃perimental results show that the average precision of the improved model is 83.7%,which is 7.4%higher than that of the DETR model,and the recall rate is increased by 10.9%.The proposed improved model can effectively detect as⁃phalt pavement damage,which can provide a reference for the detection of asphalt pavement damage in aerial images.