摘要
针对现阶段织物异常检测方法检测准确率低的问题,提出了一种改进的嵌入向量相似性的织物异常检测方法。鉴于无异常织物图像易获取,该方法不依赖负样本。通过Wide ResNet-50提取多层级特征作为嵌入向量;对不同层级的嵌入向量分别建立多元高斯分布,避免特征拼接对微小异常检测的影响;利用余弦相似性弥补马氏距离度量的局限性,提升计算异常分数的精确性,增强纹理类异常检测能力;根据不同异常特征的差异,通过SENet注意力机制对多层级异常分数图分配权重,提高异常检测准确率。MVTec数据集和AITEX织物异常数据集的实验结果表明,改进后的异常检测方法对12种织物异常的平均检测准确率为91.9%,比原始方法提升了2.6%,对复杂多样织物的异常检测的综合性能更好,且预测掩码和分割的织物异常更精确。
Aiming at the problem of low detection accuracy of current fabric anomaly detection methods,an improved fabric anomaly detection method based on embedding vector similarity was proposed.The method does not rely on negative samples in view of the easy acquisition of unabnormal fabric images.Multi-level features were extracted by Wide ResNet-50 as embedding vectors.Multivariate Gaussian distribution was established for embedding vectors of different levels to avoid the influence of feature splicing on small anomaly detection.The cosine similarity was used to make up for the limitation of Mahala Nobis distance measurement,improve the accuracy of anomaly calculation and enhance the detection ability of texture anomaly.According to the differences of different abnormal features,the SENet attention mechanism was used to assign weights to the multi-level anomaly fraction graph to improve the accuracy of anomaly detection.The experimental results of MVTec data set and AITEX fabric anomaly data set show that the average accuracy of the improved abnormal detection method for 12 kinds of fabric anomalies is 91.9%,which is 2.6%higher than the original method.The comprehensive performance of complex and diverse fabric anomaly detection is better,and the prediction of mask and split fabric anomaly is more accurate.
作者
姜金涛
丁坤
王志花
严向华
宋雅静
陈从平
JIANG Jintao;DING Kun;WANG Zhihua;YAN Xianghua;SONG Yajing;CHEN Congping(Inner Mongolia Zhicheng IOT Co.,Ltd,Ulanqab,Inner Mongolia 012001,China;College of Mechanical Engineering and Rail Transit,Changzhou University,Changzhou,Jiangsu 213164,China)
出处
《毛纺科技》
CAS
北大核心
2023年第6期73-80,共8页
Wool Textile Journal
基金
国家重点研发计划重点专项(2018YFC1903101)
科技部“科技助力经济”重点专项(SQ2020YFF0406540)。
关键词
异常检测
特征提取
距离度量
注意力机制
anomaly detection
feature extraction
distance metric
attention mechanism