SGDB: Semantic-Geometric Decoupling and Evidence Binding for Logical Anomaly Detection

Chenrui Ji, Jiafan Lu, Zilu Luo, Feng Wan, Qiuhui Chen, Weimin ZhongarXiv 2026 · arXiv

TL;DR

SGDB (Semantic–Geometric Decoupling and Evidence Binding) is a new framework for logical anomaly detection that separates object identification from geometric measurement before binding them back together for rule-based verification. By decoupling these tasks, it achieves state-of-the-art performance on the MVTec LOCO and VID-AD benchmarks while providing human-readable, evidence-backed explanations for every alarm.

What Problem Does This Paper Address?

In industrial inspection, "structural" anomalies (like a scratch on a surface) are relatively easy to spot. "Logical" anomalies are much trickier: they occur when the components themselves look normal, but their configuration violates a rule—for example, a bag containing two screws instead of three, or a rope that is 20% too long.

Current methods often struggle because they conflate appearance and geometry. A neural network might correctly identify a screw but produce a slightly inaccurate segmentation mask, leading to a "noisy" length measurement that triggers a false alarm. Conversely, a "clean" fragment of a broken object might be measured accurately but lead to an incorrect count. The authors argue that we need to decouple what an object is from how big it is, and then explicitly bind that evidence together to check against task-specific constraints.

Methods

The SGDB framework consists of three main stages:

  1. Fact Construction (Decoupling): The system uses a shared visual frontend but splits the processing into two paths. A semantic decoder (using DETR-style queries) learns to identify and count objects. Simultaneously, a witness reader extracts initial geometric data (like axes and bounding boxes).
  2. Evidence Binding: This is the "glue." A partial one-to-one association algorithm (using the Hungarian method) matches semantic identities to geometric witnesses based on class and spatial proximity. Crucially, if an object is identified but no high-quality geometry is found, the object is still counted, but its geometry is marked as "unavailable."
  3. Support-Certified Readout: For objects where geometry is critical (like screws or tapes), the system performs a refinement step. It evaluates the quality of mask-derived measurements against a "quality floor." It only updates the measurement if the new axis is high-quality and belongs to a complete, distinct component.
The SGDB framework: Fact construction, constraint verification, and grounded explanation.

Finally, an Execution-Grounded Explanation module traces the decision. It identifies a "minimum-size sufficient certificate"—the smallest set of rules that were violated to trigger the alarm. A frozen Large Language Model (Qwen3-VL) then translates these violations into a natural language report that cites specific numeric evidence.

Datasets

The authors evaluated SGDB on two major benchmarks:

  • MVTec LOCO AD: The standard for logical anomaly detection, featuring categories like "Screw bag" and "Breakfast box."
  • VID-AD: A newer, more challenging dataset containing 50 tasks across various scenarios (e.g., Ropes, Cookies, Tools) under difficult capture conditions like blur and low light.

Key Insights and Ideas

  • Decoupling is Defensive: By separating semantics from geometry, the system prevents errors in one domain from automatically corrupting the other. A bad mask won't necessarily break the object count.
  • The "Aha!" of Evidence Binding: Figure 1 illustrates the core motivation: two scenes can have identical component counts but different geometric validity. Binding ensures that the correct measurement is checked against the correct object's specific bounds.
Motivation: identical counts can still be anomalous due to geometric violations.
  • Explainability as a Certificate: Instead of just outputting a heatmap, SGDB provides a "certificate." If a bag is flagged, the system can say exactly why: "Screw O4 length is 428.2 px, which exceeds the upper bound of 363.1 px."

Key Results

SGDB sets a new high bar for logical anomaly detection:

  • MVTec LOCO: Achieved 98.19% AUROC and 94.52% F1-score on logical categories, significantly outperforming previous methods like SALAD (96.42% AUROC) and LogiCo (94.49% AUROC).
  • VID-AD: Achieved 93.56% AUROC across all 50 tasks, showing remarkable robustness to capture distractions like blur.
  • Qualitative Success: The generated reports (shown in Figure 6) successfully connect visual evidence to logical claims, making the system's "reasoning" transparent to human operators.
Examples of execution-grounded reports showing numeric evidence and LLM-generated claims.

Limitations and Future Work

  • Supervision Requirements: The method requires component-level annotations (boxes, masks, axes) for the "normal" training data, which can be labor-intensive to produce.
  • Manual Rule Specification: Task requirements (e.g., "there must be exactly two bolts") are currently manually specified. Automating the discovery of these logical rules from normal data alone is a promising future direction.
  • Unfamiliar Scenes: The authors note that extending the system to handle completely unfamiliar scenes with reduced supervision remains a challenge.

Summary updated

Abstract

Logical anomaly detection requires recognizing scene components and measuring whether their configuration satisfies task constraints. Correct identities can coexist with inaccurate geometry; clean mask fragments can yield misleading lengths. We propose Semantic-Geometric Decoupling and Evidence Binding (SGDB). A semantic decoder learns object identities from component-annotated normal sources and factual edits; a separate reader supplies final measurements. Partial one-to-one association binds geometric witnesses to identities while retaining unmatched objects in the count. A support-certified reader selects mask-derived axes using line quality and complete component ownership. Explicit task requirements and normal-data references define executable checks, with decision thresholds fixed by independent normal calibration. Execution-grounded explanation traces rule contributions to source objects and measurements, extracting compact sufficient alarm certificates. A frozen language model expresses selected outcomes as cited claims, with numeric evidence retained by the executor. Reports preserve observed deviations and unavailable evidence. SGDB achieves 98.19% AUROC and 94.52% fixed-threshold F1 on five MVTec LOCO logical categories, and 93.56% AUROC and 92.76% F1 across all 50 VID-AD tasks. Frozen-output controls evaluate geometric readout and explanatory sufficiency and compactness.
BibTeX
@misc{ji2026sgdb,
  author = {Chenrui Ji and Jiafan Lu and Zilu Luo and Feng Wan and Qiuhui Chen and Weimin Zhong},
  title = {{SGDB: Semantic-Geometric Decoupling and Evidence Binding for Logical Anomaly Detection}},
  year = {2026},
  eprint = {2610.34897},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.34897},
}
Figure