基于目标检测的工程施工物料库存智能盘点方法

Intelligent inventory counting method for construction materials based on object detection

摘要

提高工程施工物料库存智能盘点的自动化程度对于提升工程施工效率和工程质量具有重要意义。旨在通过数字图像处理技术，提出一种基于改进YOLOv5算法的钢筋自动计数方法。改进YOLOv5的网络结构，加入小目标检测层得到更大的特征图，同时使用加权双向特征金字塔融合多尺度特征图，保留高层语义信息，相比原YOLOv5x模型参数量减少，模型的鲁棒性和推理速度有所提升，可以更好地识别边缘和被遮挡的端面小的钢筋。另外，针对钢筋数据集过少且大尺度样本和小尺度样本分布不均匀的问题，使用几何图像增强、添加噪声、调整亮度等数据增强方法进行数据集扩充。在试验中，将融合小目标检测层和加权双向特征金字塔的YOLOv5模型与原模型在符合实际工况的测试集上进行对比试验。试验结果显示，该模型在检测精度和推理速度上均有提高，平均精度均值达97.00%,相较YOLOv5x提升了1.30%,每秒帧数达59.79,较YOLOv5x提升了6.88。同时，F1-score达到96.00%,表明模型已满足实际工况要求，可以进行工程部署。本研究成果为工程钢筋计数管理提供了智能化的技术方法。

Abstract

Enhancing the automation level of inventory checks for construction materials is of great significance for improving construction efficiency and project quality. An automatic rebar counting method was proposed based on the improved YOLOv5 algorithm using digital image processing technology. The YOLOv5 network structure is improved by adding a small-object detection layer to obtain larger feature maps, while a weighted bidirectional feature pyramid is used to fuse multi-scale feature maps, retaining high-level semantic information. Compared to the original YOLOv5x model, the number of parameters is reduced, and the model's robustness and inference speed are enhanced, enabling better detection of small rebars at the edges or those partially occluded. Additionally, to address the issue of limited rebar datasets and the uneven distribution of large-scale and small-scale samples, data augmentation method such as geometric image enhancement, noise addition, and brightness adjustment are used to expand the dataset. In the experiments, the YOLOv5 model with the integrated small-object detection layer and weighted bidirectional feature pyramid is compared with the original model on a test set reflecting actual working conditions. The experimental result show that the improved model achieves better detection accuracy and inference speed, with an average precision of 97.00%, an improvement of 1.30% over YOLOv5x, and a frame rate of 59.79 frames per second, which is 6.88% improvement over YOLOv5x. Meanwhile, the F1-score reaches 96.00%, indicating that the model meets the practical requirements and is ready for deployment in engineering projects. The research findings provide an intelligent technical approach for rebar counting and management in construction projects.

关键词

工程施工物资 / 钢筋识别 / 小目标检测 / YOLOv5

Key words

engineering construction materials / rebar recognition / small-object detection / YOLOv5

Author summay

吕沅庚(1999—), 男, 博士研究生, 主要从事水利工程施工文本智能挖掘研究。E-mail: [e-mail]

基金资助

中国长江三峡集团有限公司科研项目资助(202103551)