摘要
为了减少缺陷检测中的冗余检测,提出基于双动态头Sparse R-CNN的缺陷检测算法,2个动态头的责任不同:第1个负责不同尺度和空间的特征提取,第2个负责匹配可学习的提议特征。为了更好地提取图像细节信息,改进特征金字塔(FPN)为特征金字塔网格(FPG),并且与第1个动态头相结合进行特征提取。其次,提出了交流注意力来改进检测阶段的多头自注意力模块,减少随着迭代注意力图相似导致建模能力下降的问题。最后,改进边框回归损失函数GIoU为Alpha-CIoU,加速收敛并提升检测的精度。实验结果表明:算法在晶圆和热轧钢2种表面缺陷数据集上都取得很好效果,平均精度分别为94.3%和88.1%。
A defect detection algorithm based on dual dynamic head Sparse R-CNN was proposed to reduce redundant detection in defect detection.The responsibilities of the two dynamic heads were different,the first was responsible for feature extraction in different scales and spaces,and the second was responsible for matching learnable proposed features.The feature pyramid network(FPN)was improved to the feature pyramid grid(FPG)and combined with the first dynamic head for feature extraction.Secondly,speaking-head attention(SHA)was proposed to improve the multi-head self-attention module in the detection stage to reduce the decline of modeling ability with the similarity of the iterative attention graph.Finally,the loss function GIoU of bounding box regression was improved by Alpha-CIoU,which accelerated convergence and improves the accuracy of detection.The experimental results show that the algorithm has achieved great results on the surface defect data sets of the wafer and hot rolled steel,and the average accuracy is 94.3%and 88.1%respectively.
作者
郑亚睿
蒋三新
ZHENG Ya-rui;JIANG San-xin(School of Electronic and Information Engineering,Shanghai Electric Power University,Shanghai 201306,China)
出处
《仪表技术与传感器》
CSCD
北大核心
2023年第5期97-105,111,共10页
Instrument Technique and Sensor
关键词
表面缺陷检测
动态头
稀疏预测
注意力机制
标签匹配
端到端预测
surface defect detection
dynamic head
sparse prediction
attention mechanism
label matching
end-to-end prediction