基于感知哈希算法的特征融合玻璃瓶缺陷检测  

Feature Fusion Glass Bottle Defect Detection Based on Perceptual Hash Algorithm

作　　者：傅莉 吉宏轩 张宇峰[2] 任艳[3] FU Li;JI Hongxuan;ZHANG Yufeng;REN Yan(College of Automation,Shenyang Aerospace University,Shenyang 110136,China;College of Physical Science and Technology,Bohai University,Jinzhou 121013,China;College of Artificial Intelligence,Shenyang Aerospace University,Shenyang 110136,China)

机构地区：[1]沈阳航空航天大学自动化学院,辽宁沈阳110136 [2]渤海大学物理科学与技术学院,辽宁锦州121013 [3]沈阳航空航天大学人工智能学院,辽宁沈阳110136

出　　处：《无线电工程》2024年第1期55-62,共8页Radio Engineering

基　　金：国家自然科学基金(61602321)。

摘　　要：特征提取作为玻璃瓶缺陷检测任务中至关重要的一环,特征集中丰富的特征信息将直接影响缺陷检测的准确率。传统的单一特征提取算法提取的特征信息往往过于单一,使得最终的检测准确率偏低。针对上述问题,提出了方向梯度直方图(Histogram of Oriented Gradients,HOG)特征与尺度不变特征变换(Scale Invariant Feature Transform,SIFT)特征融合的特征提取算法。针对不同缺陷边缘提取轮廓不够准确的问题,提出了基于感知哈希算法(Perceptual Hash Algorithm,PHA)的边缘检测算子选择方法。通过支持向量机(Support Vector Machine,SVM)进行训练和验证。实验结果表明,提出的边缘检测算子选择方法可以针对不同缺陷选择最适合的边缘检测算子,特征融合算法的瓶身缺陷检测平均准确率可达88.7%。较单一的HOG特征提取算法提升了7.99%,较单一的SIFT特征提取算法提升了2.97%。Feature extraction is a crucial step in glass bottle defect detection task.The rich feature information in feature set will directly affect the accuracy of defect detection.However,the feature information extracted by the traditional single feature extraction algorithm is often too simple,leading to a low accuracy of the final detection.To solve these problems,a feature extraction algorithm based on the fusion of Histogram of Oriented Gradients(HOG)feature and Scale Invarient Feature Transform(SIFT)feature is proposed.To address the problem that contour extraction from different defect edges is not accurate enough,an edge detection operator selection method based on Perceptual Hash Algorithm(PHA)is proposed.Support Vector Machine(SVM)is used for training and verification.Experimental results show that the edge detection operator selection method proposed can select the most suitable edge detection operator for different defects,and the average accuracy of the feature fusion algorithm can reach 88.7%.Compared with the single HOG feature extraction algorithm,it is improved by 7.99%,and compared with the single SIFT feature extraction algorithm,it is improved by 2.97%.

关 键 词：缺陷检测 方向梯度直方图特征 SIFT特征 支持向量机 感知哈希算法 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]