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。
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 anomaly 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.
作者
王纪康
赵旭俊
WANG Ji-kang;ZHAO Xu-jun(School of Computer Science&Technology,Taiyuan University of Science and Technology,Taiyuan 030024,China)
出处
《计算机技术与发展》
2024年第5期149-156,共8页
Computer Technology and Development
基金
山西省基础研究计划资助项目(202303021221142)。
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
image anomaly detection
residual networks
knowledge distillation
attention mechanisms
transfer learning
