Abstract: To address the problem of the lack of high precision detection algorithms for material fine grain aspects in the intelligent shoveling process of loaders,an improved material fine grain object detection algorithm based on YOLOv5 was proposed.This method primarily employed an attention mechanism to enhance the model's capability in extracting fine-grained features and detecting low-quality data.To further leverage attention for optimizing network performance,a bilinear attention mechanism was introduced.The optimal embedding scheme was investigated,and the concept of soft thresholding was integrated with the bilinear attention mechanism,aiming to mitigate the impact of low-quality data on the model's detection accuracy.Experimental results demonstrated that compared to the original YOLOv5,the network improved with bilinear attention mechanism achieved a 6.0% increase in mAP@0.5 to 93.2% on high-quality samples,with a Frames Per Second (FPS) of 52.6.After embedding the soft threshold,the network's mAP@0.5 on low-quality samples was improved by 9.9% to 90.2%,with an FPS of 50.0,meeting the requirements for accuracy and real-time performance in the intelligent shovel loading process of loaders.

Key words: YOLOv5, smart shovel, material identification, fine-grained, attention mechanism, object detection

摘要： 针对装载机智能铲装过程中缺少对物料细粒度方面的高精度检测算法问题,提出基于YOLOv5改进的物料细粒度目标检测算法。该方法主要利用注意力机制提高模型对细粒度特征的提取能力和对低质量数据的检测能力。为进一步利用注意力优化网络性能,提出双线性注意力机制,研究最佳嵌入方案并将软阈值思想与双线性注意力机制结合,以达到缓解低质量数据对模型检测精度影响的目的。实验结果表明,相较于原YOLOv5,双线性注意力机制改进后的网络在高质量样本上的mAP@0.5为93.2%,提高6.0%,每秒检测帧数(FPS)为52.6;嵌入软阈值后,网络在低质量样本上的mAP@0.5为90.2%,提高9.9%,FPS为50.0,满足装载机智能铲装过程对算法检测精度和实时性的要求。

关键词: YOLOv5算法, 智能铲装, 物料识别, 细粒度, 注意力机制, 目标检测