摘要
针对当前的路面缺陷检测算法在精度和效率方面存在一定的问题,该文在YOLOv8n的基础上进行改进,提出了双分支多尺度特征下的路面缺陷检测算法YOLOv8-PFMD。首先,使用部分可变形卷积(P-DCNv3)替换常规卷积,以在提高模型特征提取能力的同时增强其对不同缺陷形变的适应能力;其次,在C2f模块中采用了更高效的Faster_RFE_Bottleneck模块,结合Pconv和RFE结构充分利用特征映射中感受野的优势,以进一步降低模型计算量;然后,在坐标注意力的基础上提出多尺度双分支坐标注意力(MDCA),通过扩展双分支的拆分融合,从而在减少模型参数的同时提高模型特征表达能力;最后,将YOLOv8n两个检测头的卷积融合成深度可分离卷积(DSConv),使模型的参数量大幅下降。实验结果表明,在RDD2022数据集和Road Damage数据集上,改进的算法与原算法相比,mAP50分别提升了8.4%、7.3%,参数量和计算量分别降低了16.7%、20.7%。在RDD2022数据集上,算法在mAP50和F1分数方面,相较于Faster-RCNN、YOLOv7等主流目标检测算法也取得了提升的效果。
Aiming at the problem of the current road defect detection algorithm for improvement in terms of accuracy and efficiency,a pavement defect detection algorithm YOLOv8-PFMD under dual-branch multi-scale features was proposed in this paper.First,partially deformable convolution(P-DCNv3)was used to replace conventional convolution to improve the feature extraction capability of the model while enhancing its adaptability to different defect deformations;secondly,the more efficient Faster_RFE_Bottleneck module was used in the C2f module,combining the Pconv and RFE structures to make full use of the advantages of the receptive field in the feature map to further reduce the amount of model calculations;then,based on the coordinate attention,a multi-scale dual-branch coordinate attention(MDCA),by expanding the split fusion of dual branches,thereby reducing model parameters and improving the model feature expression ability;finally,the convolution of the two detection heads of YOLOv8n was fused into a depth separable convolution(DSConv),making the model number of parameters was significantly reduced.Experimental results showed that on the RDD2022 data set and Road Damage data set,compared with the original algorithm,the mAP50 of the improved algorithm increased by 8.4%and 7.3%respectively,and the amount of parameters and calculation amount were reduced by 16.7%and 20.7%respectively.On the RDD2022 data set,the algorithm achieved improved results compared to mainstream target detection algorithms such as Faster-RCNN and YOLOv7 in terms of mAP50 and F1 scores.
作者
付景泽
吕伏
FU Jingze;LYU Fu(School of Software,Liaoning Technical University,Huludao,Liaoning 125105,China;Department of Basic Teching,Liaoning Technical University,Huludao,Liaoning 125105,China)
出处
《测绘科学》
CSCD
北大核心
2024年第10期36-49,共14页
Science of Surveying and Mapping
基金
国家自然科学青年基金项目(51904144)
国家自然科学基金面上项目(51874166,52274206,51974145)。