摘要
我国是工业铝型材制造大国,铝型材生产质量检测意义重大。针对传统的人工目测等方式检测效率低下,稳定性相对较弱;单一YOLOv3方法特征提取不突出,检测精度有限等问题,提出一种基于图像融合与YOLOv3的铝型材表面缺陷检测方法。首先利用图像增强、空域滤波的方法对原始图像进行预处理得到处理图像;然后借鉴SLAM中特征提取与匹配的思想对原始图像和处理图像进行特征提取与匹配;之后进行图像融合得到最终的处理后图像;再通过K-means算法聚类和调参优化,最后利用单阶段物体检测模型YOLOv3对铝型材表面缺陷进行检测。通过一个end-to-end的全卷积神经网络完成从原始图像的输入到Bounding box和box中物体类别与置信度的输出。实验结果表明,此图像融合与YOLOv3的方法对表面缺陷分类检出的平均成功率为98.33%,比单一YOLOv3方法提高了3.75个百分点;验证集mAP值为88.81%,提高了4.18个百分点,具有更强的特征提取能力和泛化能力,能精确检测表面缺陷,进行分类和定位。
China is a large country in the manufacture of aluminum profiles,and the quality inspection of aluminum profiles is of great significance.Aiming at the problems of low detection efficiency and relatively weak stability of traditional manual visual inspection methods,the feature extraction of the single YOLOv3 method is not prominent,and the detection accuracy is limited,this paper proposes a method for detecting surface defects of aluminum profiles based on image fusion and YOLOv3.First,image enhancement and spatial filtering methods are used to preprocess the original image to obtain a processed image;then the feature extraction and matching ideas in SLAM are used for reference to extract and match the original image and the processed image;then image fusion is performed to obtain the final processed image;subsequently,the K-means algorithm is used for clustering and tuning parameters optimization.Finally,a single-stage object detection model YOLOv3 is used to detect the surface defects of aluminum profiles.An end-to-end full convolutional neural network is used to complete the input from the original image to the output of the object categories and confidence in the Bounding box and box.The experimental results show that the average success rate of surface defect classification detection by this method is 98.33%,which is 3.75 percent points higher than the single YOLOv3 method;the mAP value of the validation set is 88.81%,which has an increase of 4.18 percent points.It has stronger feature extraction and generalization capabilities,and can accurately detect surface defects,perform classification and localization.
作者
张磊
郎贤礼
王乐
ZHANG Lei;LANG Xian-li;WANG Le(School of Instrument Science and Optoelectronic Engineering,Hefei University of Technology,Hefei 230009,China)
出处
《计算机与现代化》
2020年第11期8-15,共8页
Computer and Modernization
基金
国家自然科学基金资助项目(51505120)
中央高校基础科研基金资助项目(JZ2019HGTB0082)
安徽省横向项目(W2019JSKF0133)。
