基于多尺度特征模仿的无监督异常定位算法

An unsupervised anomaly localization method based on multi-scale feature imitation

无监督异常定位任务面临缺少异常样本训练,需要检测多种类别异常和处理多种异常区域面积占比的挑战。针对这些问题,提出一种分离式教师-学生特征模仿网络结构和一种结合图像金字塔和特征金字塔的多尺度处理策略,并提出了一种基于梯度下降优化的网络模块重要性搜索的方法以简化网络结构。实验结果表明,在真实工业产品检测数据集上,所提出的算法比同时期的特征建模异常定位方法表现更好,多尺度策略对比基准方法能有效实现效果改善。

Unsupervised anomaly localization tasks face the challenge of lacking anomaly sample for training,of the need to detect multiple types of anomalies and handle the various proportions of multiple abnormal regions.Targeting on these problems,a separate teacher-student feature imitation network and a multi-scale processing strategy combining image pyramid and feature pyramid are proposed.A method of importance search for network modules based on gradient descent optimization is proposed to simplify network structure.The experimental results show that on the real industrial product detection dataset,the proposed algorithm performs better than the feature modeling anomaly location method of the same period.The multi-scale strategy can effectively improve the performance compared with the benchmark method.