Measurement-Driven LSAD: An Interpretable Lightweight Distillation Framework for Edge-Device Industrial Logical Anomaly Detection

In industrial anomaly detection, large-scale models are widely used for texture analysis and defect recognition. However, lightweight deployment still faces challenges in achieving a balance between measurement accuracy and processing speed. Existing knowledge distillation models often suffer from degraded representation ability because of excessive simplification in student networks. Segmentation networks for anomaly detection also lack lightweight designs, which makes it difficult to meet the real-time requirements of resource-constrained edge devices. To address these issues, this paper proposes a Lightweight Hybrid Convolutional \& Segmentation Anomaly Detection model (LSAD). The model is built on a unified framework that integrates a Lightweight Hybrid Convolutional Distillation module (LHCD) with a lightweight segmentor (LiteSegmentor). LHCD uses a hierarchical depthwise separable teacher-student structure to form a ternary network. The intermediate module (TA) enhances the representation of the student model through a multi-scale feature pyramid. A Hybrid Convolutional Feature Extractor (HFN) and a bidirectional distillation mechanism in TA provide acceleration and parameter compression while preserving feature representation. LiteSegmentor reconstructs the Atrous Spatial Pyramid Pooling (ASPP) module with channel distillation and structural reparameterization. This design achieves a 98% reduction in parameters and a 42% acceleration in inference compared with the baseline. Experiments show that LSAD significantly improves anomaly measurement accuracy while maintaining fast inference and efficient memory use. The proposed model offers an effective solution for real-time anomaly detection and measurement on industrial edge devices.