Abstract: Abstract: In modern assembly lines, industrial anomaly detection was challenged by dual tasks: ensuring the structural integrity of single-component surfaces while verifying logical assembly relationships among multiple components. Logical anomalies were defined as global defects violating product assembly constraints. Existing unsupervised anomaly detection methods were observed to lack an optimal balance between accuracy and speed for logical anomaly detection. To address this, LogicRD was proposed to enhance logical anomaly detection capability while preserving the local anomaly detection performance and low latency of reverse distillation (RD). First, a Global Semantic Enhancement (GSE) module was designed to transform multi-scale features from the teacher network into multi-level global semantic representations containing spatial-logical relationships through cross-hierarchical feature fusion. Second, a Feature Compensation Pathway (FCP) was introduced across network layers to alleviate fine-grained information loss in global semantics, thereby improving reconstruction accuracy of the student network. A Normal Information Inhibition (NII) gate mechanism was simultaneously implemented to eliminate potential anomaly leakage risks. Finally, considering that logical anomalies could not be effectively expressed through anomaly maps, Mahalanobis distance was employed to directly measure feature deviations. Logical anomalies were evaluated jointly through this metric and anomaly maps, establishing a dual-branch evaluation strategy for comprehensive industrial anomaly assessment. On the MVTec LOCO benchmark, LogicRD achieved 83.8% image-level AUROC with 55 FPS inference speed, showing 7 percentage-point improvement over baseline RD. Comparable performance was attained with DSKD (a dual-branch logical anomaly detection model based on RD), while achieving 1.7× faster inference speed.

Key words: unsupervised deep learning, logic anomaly detection, reverse distillation, multi-layered global semantics, mahalanobis distance

摘要： 摘 要: 现代装配生产线中，工业异常检测面临双重任务挑战：在确保单组件表面结构完整性的同时，还需验证多组件间的装配逻辑关系。逻辑异常是指违反产品装配逻辑约束的全局性缺陷。现有的无监督异常检测方法在逻辑异常检测上的准确度与速度上并未取得一个良好的权衡。为此，提出了LogicRD，在保持反向蒸馏（RD）的局部异常检测能力和低延迟的情况下，扩展其逻辑异常检测的能力。首先，全局语义增强模块（GSE）通过特征跨层次融合策略，将教师网络的多尺度特征转化为包含空间逻辑关系的多层次全局语义表征。其次，增加跨层级特征补偿通路（FCP），缓解全局语义中的细粒度信息丢失问题，提升学生重建精度，并结合正常信息门控机制（NII）以消除潜在异常泄漏的风险。最后，存在异常图难以表达的逻辑异常，利用马氏距离对特征的偏离程度直接评估并和异常图对逻辑异常进行联合评价，双分支评价策略可以对工业异常进行全面综合的评价。在MVTec LOCO基准测试中，LogicRD的图像级接收者特性曲线下的面积（Image-AUROC）达到了83.8%，推理速度达到了每秒55帧，较基准模型RD提升了7个百分点，达到了基于RD的双分支逻辑异常检测模型DSKD相近的性能，但推理速度是DSKD的1.7倍。

关键词: 无监督深度学习, 逻辑异常检测, 反向蒸馏, 多层次全局语义, 马氏距离