[1]王纪康,赵旭俊.基于知识蒸馏的图像异常检测方法[J].计算机技术与发展,2024,34(05):149-156.[doi:10.20165/j.cnki.ISSN1673-629X.2024.0053]
　WANG Ji-kang,ZHAO Xu-jun.An Image Anomaly Detection Method Based on Knowledge Distillation[J].,2024,34(05):149-156.[doi:10.20165/j.cnki.ISSN1673-629X.2024.0053]
点击复制

基于知识蒸馏的图像异常检测方法()

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:
34
期数:
2024年05期
页码:
149-156
栏目:
人工智能
出版日期:
2024-05-10

文章信息/Info

Title:
An Image Anomaly Detection Method Based on Knowledge Distillation
文章编号:
1673-629X(2024)05-0149-08
作者:
王纪康; 赵旭俊
太原科技大学 计算机科学与技术学院,山西 太原 030024
Author(s):
WANG Ji-kang; ZHAO Xu-jun
School of Computer Science & Technology,Taiyuan University of Science and Technology,Taiyuan 030024,China
关键词:
图像异常检测; 残差网络; 知识蒸馏; 注意力机制; 迁移学习
Keywords:
image anomaly detection; residual networks; knowledge distillation; attention mechanisms; transfer learning
分类号:
TP391
DOI:
10.20165/j.cnki.ISSN1673-629X.2024.0053
摘要:
图像异常检测中模型的浅层架构对细微差异有较弱的检测能力,寻找有效的特征表示来区分正负样本是一个挑战。 为此,提出了一种新的基于知识蒸馏的图像异常检测方法。 该方法提出一种新的知识蒸馏框架,由 T-S 模型和单类嵌入模块组成,通过迁移学习泛化新异常。 首先,高容量的 wide_resnet50_2 网络作为教师网络,通过单类嵌入模块在最低层次将多尺度特征聚合,保留普遍性和空间分辨率,增强了蒸馏模型对异常的表示能力。 其次,嵌入注意力机制的工作上,在保持网络结构完整性的同时,为预训练参数的有效利用提供了新的视角,提高了模型的性能。 最后,提出了一种新的异常表示方法,计算每对张量的余弦相似损失,累计多尺度异常得到异常分数图。 实验结果表明,该方法在 MVTec 数据集的纹理和物体类别上,平均 AUC 值分别达到了 97. 8% 和 95. 5% ,对图像中的细微异常具有优秀的检测能力。
Abstract:
The shallow architecture of the model in image anomaly detection has a weak ability to detect subtle differences,and it is a challenge to find effective feature representations to distinguish between positive and negative samples. To solve this problem,a new image anomaly detection method based on knowledge distillation was proposed. This method proposes a new knowledge distillation framework,which consists of a T - S model and a single - class embedding module, and generalizes new anomalies through transfer learning. Firstly,the high-capacity wide_resnet50_2 network as a teacher network aggregates multi-scale features at the lowest level through a single-class embedding module,retains the universality and spatial resolution,and enhances the ability of the distillation model to represent anomalies. Secondly,the work of embedding attention mechanism provides a new perspective for the effective use of pre-trained parameters and improves the performance of the model while maintaining the integrity of the network structure. Finally,a new a-nomaly representation method is proposed,which calculates the cosine similarity loss of each pair of tensors,and accumulates multi-scale anomalies to obtain the anomaly score graph. Experimental results show that the proposed method achieves an average AUC value of 97.8% and 95.5% on the texture and object class of the MVTec dataset, respectively, and has excellent detection ability for subtle anomalies in the image.

相似文献/References:

[1]赵嘉兴,王夏黎,王丽红,等.多尺度密集时序卷积网络的单幅图像去雨方法[J].计算机技术与发展,2020,30(05):115.[doi:10. 3969 / j. issn. 1673-629X. 2020. 05. 022]
　ZHAO Jia-xing,WANG Xia-li,WANG Li-hong,et al.Single Image De-raining Method for Multi-scale Dense Temporal Convolutional Networks[J].,2020,30(05):115.[doi:10. 3969 / j. issn. 1673-629X. 2020. 05. 022]
[2]李 栋,张 蕾*,郭茂祖,等.基于时空卷积残差网络的空气质量预测[J].计算机技术与发展,2020,30(06):124.[doi:10. 3969 / j. issn. 1673-629X. 2020. 06. 024]
　LI Dong,ZHANG Lei *,GUO Mao-zu,et al.Air Quality Prediction Based on Spatio-temporal Convolution Residual Network[J].,2020,30(05):124.[doi:10. 3969 / j. issn. 1673-629X. 2020. 06. 024]
[3]周传华,吴幸运,李 鸣.基于 WGAN 单帧人脸图像超分辨率算法[J].计算机技术与发展,2020,30(09):29.[doi:10. 3969 / j. issn. 1673-629X. 2020. 09. 006]
　ZHOU Chuan-hua,WU Xing-yun,LI Ming.Single Frame Face Images Super-resolution Algorithm Based on WGAN[J].,2020,30(05):29.[doi:10. 3969 / j. issn. 1673-629X. 2020. 09. 006]
[4]焦 亮,张太红*.基于深度学习身份证鉴别与信息检测方法研究[J].计算机技术与发展,2020,30(12):203.[doi:10. 3969 / j. issn. 1673-629X. 2020. 12. 036]
　JIAO Liang,ZHANG Tai-hong*.Research on Identity Card Identification and Information Detection Based on Deep Learning[J].,2020,30(05):203.[doi:10. 3969 / j. issn. 1673-629X. 2020. 12. 036]
[5]江佳俊,蒋 旻*,杨晓雨,等.基于注意力机制的个性化图像美学质量评估[J].计算机技术与发展,2021,31(10):56.[doi:10. 3969 / j. issn. 1673-629X. 2021. 10. 010]
　JIANG Jia-jun,JIANG Min*,YANG Xiao-yu,et al.Research on Evaluation of Personalized Image Aesthetic Quality Based on Attention Mechanism[J].,2021,31(05):56.[doi:10. 3969 / j. issn. 1673-629X. 2021. 10. 010]
[6]鲍先富,强赞霞,李丹阳,等.基于组卷积特征融合的 One-Stage 目标检测模型[J].计算机技术与发展,2021,31(11):86.[doi:10. 3969 / j. issn. 1673-629X. 2021. 11. 015]
　BAO Xian-fu,QIANG Zan-xia,LI Dan-yang,et al.One-Stage Target Detection Model Based on Group ConvolutionFeature Fusion[J].,2021,31(05):86.[doi:10. 3969 / j. issn. 1673-629X. 2021. 11. 015]
[7]谢斌红,赵金朋,张英俊.结合注意力机制的车型检测算法[J].计算机技术与发展,2021,31(12):78.[doi:10. 3969 / j. issn. 1673-629X. 2021. 12. 014]
　XIE Bin-hong,ZHAO Jin-peng,ZHANG Ying-jun.Vehicle Detection Algorithm Combined with Attention Mechanism[J].,2021,31(05):78.[doi:10. 3969 / j. issn. 1673-629X. 2021. 12. 014]
[8]姜丽莉,黄承宁.融合注意力机制改进残差网络的表情识别方法[J].计算机技术与发展,2022,32(05):42.[doi:10. 3969 / j. issn. 1673-629X. 2022. 05. 007]
　JIANG Li-li,HUANG Cheng-ning.An Expression Recognition Method Based on Fusion of Attention Mechanism and Improved Residual Network[J].,2022,32(05):42.[doi:10. 3969 / j. issn. 1673-629X. 2022. 05. 007]
[9]彭 治,刘 杨,杜永萍,等.基于迁移学习的多场景垃圾图像分类方法[J].计算机技术与发展,2022,32(05):106.[doi:10. 3969 / j. issn. 1673-629X. 2022. 05. 018]
　PENG Zhi,LIU Yang,DU Yong-ping,et al.Multi-scene Garbage Image Classification Method Based on Transfer Learning[J].,2022,32(05):106.[doi:10. 3969 / j. issn. 1673-629X. 2022. 05. 018]
[10]杨朝晨,陈佳悦,邢 可,等.基于改进的 DSSD 的小目标检测算法研究[J].计算机技术与发展,2022,32(06):63.[doi:10. 3969 / j. issn. 1673-629X. 2022. 06. 011]
　YANG Zhao-chen,CHEN Jia-yue,XING Ke,et al.Small Target Detection Algorithm Based on Improved DSSD[J].,2022,32(05):63.[doi:10. 3969 / j. issn. 1673-629X. 2022. 06. 011]

更新日期/Last Update: 2024-05-10