基于改进YOLOv5的工件识别算法

Workpiece recognition algorithm based on improved YOLOv5

摘要:

针对工业生产流水线中工件识别速度慢、精度低的问题,提出一种基于改进YOLOv5的工件识别方法,称为YOLO_Meta。对YOLOv5原有的网络架构进行了多个阶段的调整,包括利用双路注意力机制模块和深度可分离卷积对主干特征提取网络进行改进,可以更全面的提取特征；引入一种新型解耦头增强模型对各层级特征图的利用效率；利用KMeans算法计算随机锚框相似度的方法对先验框进行过滤以及加入标签平滑算法等。本文基于MS COCO和自制工件数据集进行实验并根据模型深度和宽度将模型分为大、中、小三款模型。实验结果表明,在MS COCO数据集上大、中、小三款模型对比原模型的AP分别提高了3.4%、1.8%、6.9%。在自制工件数据集上大模型对比原模型mAP提高了19.1%,F1分数提高了15.2%。本文提出的YOLO_Meta模型与原始模型相比无论是稳定性还是准确率都有很大的提升,该方法可为工件检测任务提供参考。

Abstract:

Aiming at the problems of slow workpiece recognition and low precision in industrial production lines, a workpiece recognition method based on improved YOLOv5 is proposed, called YOLO_Meta. The original network architecture of YOLOv5 has been adjusted in multiple stages, including the use of two-way attention mechanism modules and depth-separable convolutions to improve the backbone feature extraction network, which can extract features more comprehensively; introduce a new type of decoupling The head enhances the model"s ability to utilize feature maps at each level; the KMeans algorithm is used to calculate the similarity of random anchor frames to filter the prior frames and add label smoothing algorithms, etc. This paper conducts experiments based on MS COCO and self-made workpiece datasets, and divides the models into three models: large, medium and small according to the depth and width of the model. The experimental results show that the AP of the large, medium, and small models on the MS COCO dataset has increased by 3.4%, 1.8%, and 6.9%, respectively, compared with the original model. Compared with the original model, the mAP of the large model on the self-made artifact dataset has increased by 19.1%, and the F1 score has increased by 15.2%. Compared with the original model, the YOLO_Meta model proposed in this paper has greatly improved both in terms of stability and accuracy. This method can provide a reference for artifact detection tasks.