基于改进SSD的工件表面缺陷检测
Defect Detection of Workpiece Surface Based on Improved SSD
刘艳菊 1王秋霁 1张惠玉 1刘彦忠 1赵开峰1
作者信息
- 1. 齐齐哈尔大学 计算机与控制工程学院, 黑龙江 齐齐哈尔 161006
- 折叠
摘要
工件的表面缺陷不仅影响外观而且直接影响产品的质量、寿命和性能,因此对工件进行实时表面缺陷检测很有必要.针对当前SSD算法不利于小目标检测易导致误检的情况,提出了一种基于单阶段多层检测器的改进SSD自动检测方法.采用了以ResNet替换SSD中原始的VGGNet的方法,研究了小目标检测的问题;采用了对深层特征进行反卷积且将深层特征与浅层特征融合的方法,研究了语义信息不足易误检的问题.结果表明,该方法较原SSD模型在工件的表面缺陷检测上mAP值提高了约4.6%,从而认为本方法可用于工件表面缺陷的实时自动检测.
Abstract
The surface defects of the workpiece not only affect the appearance,but also affect the quality,life and performance of the product directly.Therefore it is necessary to carry out real-time surface defects detection of the workpiece.Aiming at the problems that current SSD algorithm was not suitable for small target detection,which was easy to result and error,an improved SSD automatic detection method based on single-stage multi-layer detector was proposed.The original VGGNet was replaced by ResNet in SSD.The question about small target detection was researched.By using the method of deconvolution of deep features and combine of deep features and shallow features,the problem that the semantic information which was insufficient and easy to be misdetected was studied.The results show that the mAP value of surface defect detection can be improved by the proposed method about 4.6%compared with the original SSD model.The method proposed can be used for real-time automatic detection of surface defect of the workpiece.
关键词
工件表面/缺陷检测/SSD/反卷积/特征融合Key words
workpiece surface/defect detection/SSD/deconvolution/feature fusion引用本文复制引用
基金项目
国家自然科学基金青年基金(61403222)
黑龙江省科技厅面上项目(F201439)
黑龙江省教育厅基本科研业务费科研项目(135309466)
出版年
2024
NETL
NSTL
万方数据