多尺度特征融合改进Faster RCNN的铝材表面缺陷辨识
Multi-Scale Feature Fusion to Improve Faster RCNN for Aluminum Surface Defect Recognition
陈法法 1刘咏 1潘瑞雪 1陈保家1
作者信息
- 1. 三峡大学水电机械设备设计与维护湖北省重点实验室,宜昌 443002
- 折叠
摘要
针对铝型材表面缺陷类型多样、缺陷特征复杂,难以自动检测的问题,设计了一种基于改进Faster RCNN的铝材表面缺陷检测模型.以感兴趣区域校准代替感兴趣区域池化,减少Faster RC-NN模型自身量化产生的缺陷定位误差;以Darknet-53 结合特征金字塔为主干网络提高对微小缺陷的提取能力;利用热重启的余弦退火策略更新模型的学习率,进一步加速模型收敛,提高模型检测精度.通过实际的铝型材表观缺陷数据进行测试,该方法对铝型材表面缺陷识别的平均准确率达到96.5%,单张图片检测时间为0.373 s.综合分析表明,所构建的多尺度特征融合改进Faster RC-NN的铝材表面缺陷辨识模型,能够达到工程界对铝型材表观缺陷进行缺陷辨识的实际应用需求.
Abstract
Aiming at the problem that it is difficult to automatically detect the surface defects of aluminum profiles with various types and complex defect features,an aluminum surface defect detection model based on improved Faster RCNN is designed in this paper.In this model,ROI Pooling is replaced by ROI Align to reduce the error of defect localization.Then Darknet-53 combined with FPN is used as the backbone net-work to improve the extraction ability of small defects.Finally,Cosine annealing is used to optimize the learning rate in model training to further accelerate model convergence and improve model detection accu-racy.In the final test,the average recognition accuracy of the method for aluminum surface defect can reach to 96.5%,and the detection time for a single image is only 0.373 s.
关键词
缺陷检测/Faster/RCNN/特征提取/余弦退火Key words
defect detection/Faster RCNN/feature extraction/cosine annealing引用本文复制引用
基金项目
国家自然科学基金(51975324)
国家大坝安全工程技术研究中心开放基金(CX2022B06)
湖北省教育厅科研项目(B2021036)
出版年
2024
NETL
NSTL
万方数据