基于YOLO的双层注意力缺陷检测算法
YOLO-Based Bi-Level Attention Defect Detection Algorithm
王素珍 1吕基岳 1葛润东 1邓成禹1
作者信息
- 1. 青岛理工大学信息与控制工程学院,青岛 266000
- 折叠
摘要
为了解决钢铁缺陷检测任务中小尺度、形态复杂、结构模糊目标导致现有算法精度低漏检率高问题,提出了基于YOLOv5s的SDD-YOLO算法.SDD-YOLO通过使用双层路由Transformer将局部特征与全局特征结合,提高对结构模糊的缺陷的检测效果;设计了全新的CSDA注意力,增强空间和通道的信息交互能力;使用NWD距离改进NMS算法,提高对小尺度目标的检测精度;设计了一种新的特征提取结构,降低梯度信息损失.使用增强后NEU-DET数据集实验后表明,SDD-YOLO算法相比YOLOv5s召回率提升了6.22%,平均精度均值提高了5.38%,提高了对多种缺陷类型的检测能力同时能够满足实时检测的需求.
Abstract
In order to solve the problem of low accuracy and high missed detection rate of existing algo-rithms caused by small scale,complex shape and fuzzy structure targets in steel defect detection tasks,the SDD-YOLO algorithm based on YOLOv5s is proposed.SDD-YOLO uses a two-layer routing Transformer to combine local features with global features to improve the detection of structurally ambiguous defects;a new CSDA attention is designed to enhance the information interaction capabilities of space and channels;NMS algorithm is improved by using NWD distance,to improve the detection accuracy of small-scale tar-gets;a new feature extraction structure is designed to reduce the loss of gradient information.Experiments using the enhanced NEU-DET data set show that the recall rate of the SDD-YOLO algorithm has increased by 6.22%compared with YOLOv5s,and the average precision has increased by 5.38%,it improves the detection ability of a variety of defects and can meet the needs of real-time detection.
关键词
Transformer/YOLOv5/钢铁缺陷检测/注意力机制/NWDKey words
Transformer/YOLOv5/steel defect detection/attention mechanism/NWD引用本文复制引用
基金项目
山东省自然科学基金(ZR2021MF024)
山东省自然科学基金(ZR2020QF101)
出版年
2024
NETL
NSTL
万方数据