| 摘 要: 图像异常检测旨在解决工业自动化质量控制中仅利用正常样本识别未知异常的核心挑战。针对现有方法异常样本合成真实性不足、特征重建计算开销大及特征选择缺乏自适应性的问题,提出一种自监督异常检测框架GSANet。该框架包含三个核心模块：首先,设计强度可控异常生成(Strength Controllable Anomaly Generation, SCAG)策略,利用去噪扩散概率模型在逆向采样过程中引入可控扰动,生成位于正常样本低概率密度区域的逼真异常图像；其次,构建异常特征选择模块(Feature Selection Module, FSM),通过自监督方式从预训练网络中自适应筛选最具判别力的特征子集；最后,引入残差选择模块(Residual Selection Module, RSM),从重建残差中筛选信息最丰富的特征以提升异常区域召回率。在MVTec-AD和VisA数据集上的实验结果表明,GSANet在图像级与像素级异常检测任务上均取得领先性能,图像级AUROC达99.65%,像素级AUROC达99.03%,显著优于现有方法,为工业异常检测提供了有效解决方案。 |
| 关键词: 图像异常检测 扩散概率模型 自监督 特征选择 残差选择 |
|
中图分类号:
文献标识码:
|
| 基金项目: 上海市白玉兰人才计划浦江项目25PJD089 |
|
| A Framework for Industrial Anomaly Detection via Controllable Anomaly Generation and Adaptive Feature Selection |
|
QIN Xiaofei, SHEN Xiaosong, MENG Yejia, YANG Lexiang
|
School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology
|
| Abstract: Image anomaly detection aims to solve the core challenge of identifying unknown anomalies using only normal samples in industrial automation quality control. To address the issues of insufficient realism in synthesized anomaly samples, high computational cost of feature reconstruction, and lack of adaptivity in feature selection in existing methods, we propose a self-supervised anomaly detection framework called GSANet. The framework consists of three core modules. First, a Strength Controllable Anomaly Generation (SCAG) strategy is designed. It utilizes a denoising diffusion probabilistic model to introduce controllable perturbations during the reverse sampling process, generating realistic anomaly images located in low-probability density regions of normal samples. Second, an anomaly Feature Selection Module (FSM) is constructed, which adaptively selects the most discriminative feature subsets from pre-trained networks in a self-supervised manner. Finally, a Residual Selection Module (RSM) is introduced to filter the most informative features from reconstruction residuals, thereby improving the recall rate of anomalous regions. Experimental results on the MVTec-AD and VisA datasets demonstrate that GSANet achieves leading performance on both image-level and pixel-level anomaly detection tasks, attaining an image-level AUROC of 99.65% and a pixel-level AUROC of 99.03%, significantly outperforming existing methods and providing an effective solution for industrial anomaly detection. |
| Keywords: Image Anomaly Detection Diffusion Probabilistic Models Self-supervised Feature Selection Residual Selection |
