计算机应用

• 人工智能与仿真 •    下一篇

结合非局部相似性的foveated选票缺陷检测

蒲杰1,官磊2   

  1. 1. 中国科学院成都计算机应用研究所
    2. 成都计算机应用研究所
  • 收稿日期:2019-11-26 修回日期:2019-12-23 发布日期:2019-12-23 出版日期:2020-05-09
  • 通讯作者: 蒲杰

Foveated ballot defect detection combined with non-local similarity

  • Received:2019-11-26 Revised:2019-12-23 Online:2019-12-23 Published:2020-05-09

摘要: 针对传统选票缺陷检测过程中图像配准的计算复杂度高、过程繁琐、对图案细节变化的鲁棒性差等问题, 提出了一种避免图像配准、基于 Patch 相似性度量的 foveated NL-means 缺陷检测算法。该算法是对传统 window NLmeans缺陷检测算法的改进,通过构建非局部相似模型,利用Patch权重和相似性关联对缺陷图像进行重构,无法重构的部分即为缺陷区域。通过 foveated NL-means 算法和 window NL-means算法的实验对比表明,前者对缺陷区域的检测效果更加显著;其次这两种缺陷检测算法AUC分别为:0. 923 5和0. 863 8(小于0. 923 5),数值积分表明前者对缺陷区域的预测更加精确,缺陷的分类性能更高;最后通过计算这两种算法的平均时间开销,可知foveated NL-means算法的时间效率相较于window NL-means算法平均提升了11. 697 1 s,因此能够高效的完成缺陷检测任务。

关键词: 选票缺陷检测, 非局部相似模型, Patch权重, window NL-means算法, foveated NL-means算法, 图像重构, 受 试者工作特征曲线, ROC曲线下的面积

Abstract: Aiming at the problems of high computational complexity,cumbersome process and poor robustness to pattern detail changes in the process of traditional ballot defect detection,a foveated NL(Non-Local)-means defect detection algorithm based on Patch similarity and without image registration was proposed. The algorithm is an improvement of the traditional window NL-means defect detection algorithm. By constructing a non-local similarity model,the defect image was reconstructed by using the Patch weight and the similarity association,and the part that could not be reconstructed was the defect area. The experimental comparison between the foveated NL-means algorithm and the window NL-means algorithm shows that the former has a more significant detection effect on defect areas. Moreover,the maximum AUCs(Area Under the ROC(Receiver Operating Characteristic)Curve)of these two defect detection algorithms are 0. 923 5 and 0. 863 8(less than 0. 923 5). Numerical integration shows that the former is more accurate in predicting defect areas and the classification performance of defects is higher. By calculating the average time costs of these two algorithms,it can be seen that the time efficiency of the foveated NL-means algorithm is 11. 697 1 s faster than the window NL-means algorithm.

Key words: ballot defect detection, Non-Local (NL) similarity model, Patch weight, window NL-means algorithm, foveated NL-means algorithm, image reconstruction, ROC(Receiver Operating Characteristic)curve, AUC (Area Under the ROC curve)

中图分类号: