【摘要】 针对目前已有目标检测算法在路面坑洼养护应用较少，且存在检测模型参数量较大、小目标容易漏检的问题提出一种改进的YOLOv5的算法。在主干(Backbone)层采用轻量化卷积GhostConv代替原有的标准卷积，减少模型参数；在颈部(Neck)层加入卷积GSConv和改进的注意力机制GSECA以及改进的双向融合模型BiFPN-m,增强特征信息提取与融合能力；将损失函数替换为EIOU Loss,提高小目标的检测精度。改进后的YOLOv5算法的mAP提高了3.1%,参数量降低了40%,为路面智能化养护提供了一种解决方案。更多还原

【Abstract】 An improved YOLOv5 algorithm was proposed to solve the problems that the existing target detection algorithms are rarely used in road surface pothole maintenance, and there are problems with large detection model parameters and small targets that are easy to miss. In the Backbone layer, lightweight convolution GhostConv was used to replace the original standard convolution, effectively reducing model parameters. In the Neck layer, convolution GSConv, improved attention mechanism GSECA and improved bidirectional fusion model BiFPN-m were added to enhance feature information extraction and fusion capabilities. The loss function was replaced with EIOU Loss to improve the detection accuracy of small targets. The mAP of the improved YOLOv5 algorithm is increased by 3.1%, and the parameter volume is reduced by 40%, providing a solution for intelligent road maintenance.更多还原