融合滑动注意力机制的钢带缺陷检测算法
Steel Strip Defect Detection Algorithm Integrating Sliding Attention Mechanism
赵文晶1
作者信息
- 1. 太原理工大学 工程训练国家级实验教学示范中心,太原 030024
- 折叠
摘要
针对工业钢带表面缺陷检测存在的小目标识别率差、检测精度低等问题,提出一种融合滑动注意力机制的钢带缺陷检测算法.首先,构建融合滑动注意力主干网络,建模局部自注意力机制全局上下文;其次,提出基于内容重组的上采样算子,通过模型感受野的提升捕获目标周围特征信息;最后,通过可自适应学习的参数引导特征融合模块,抑制模型在训练过程中由于梯度反传而导致的不一致性.工业钢带数据集NEU-DET上的实验结果表明,所提检测算法能够在牺牲较少检测速度的情况下,提升均值平均精度至83.2%.
Abstract
Focus on the issues of poor small object recognition rate and low detection accuracy in industrial steel strip surface defect detection,we propose a steel strip defect detection algorithm with a sliding attention mecha-nism.First,a backbone network is constructed that integrates sliding attention,and a local self-attention mechanism is used to model the global context.Second,an upsampling operator based on content reorganization is proposed to capture feature information around the object by expanding the receptive field of the models.Finally,the parameter-guided feature fusion module,which can be adaptively learned,suppresses the inconsistency of the model caused by gradient backpropagation during the training process.Experiment results on the industrial steel strip NEU-DET dataset demonstrate that the proposed detection algorithm can improve the average accuracy of the mean to 83.2%at the expense of less detection speed.
关键词
缺陷检测/注意力机制/钢带/自适应空间特征融合Key words
defect detection/attention mechanism/steel strip/adaptive spatial feature fusion引用本文复制引用
基金项目
国家自然科学基金(62006169)
山西省自然科学基金面上项目(202103021224056)
山西省回国留学人员科研资助项目(2021-046)
出版年
2024
NETL
NSTL
万方数据