LogicAL: Towards logical anomaly synthesis for unsupervised anomaly localization

TL;DR

This work tackles unsupervised anomaly localization with limited real anomaly data by introducing LogicAL, a framework that synthesizes both logical and structural anomalies through edge manipulation and edge-to-image translation. It combines region-aware edge editing (via SAM) with a pix2pixHD-based generator and TPS warping to produce realistic yet logically inconsistent defects, and it augments the localization model with edge reconstruction to improve pixel-level precision. Extensive experiments on MVTecLOCO, MVTecAD, VisA, and MADsim demonstrate state-of-the-art or competitive performance in both image-level anomaly detection and pixel-level localization, with ablations clarifying the contributions of region control, edge strategies, and reconstruction. The approach offers strong practical impact for industrial inspection by enhancing robustness to diverse defect types and logical constraints while maintaining efficient inference.

Abstract

Anomaly localization is a practical technology for improving industrial production line efficiency. Due to anomalies are manifold and hard to be collected, existing unsupervised researches are usually equipped with anomaly synthesis methods. However, most of them are biased towards structural defects synthesis while ignoring the underlying logical constraints. To fill the gap and boost anomaly localization performance, we propose an edge manipulation based anomaly synthesis framework, named LogicAL, that produces photo-realistic both logical and structural anomalies. We introduce a logical anomaly generation strategy that is adept at breaking logical constraints and a structural anomaly generation strategy that complements to the structural defects synthesis. We further improve the anomaly localization performance by introducing edge reconstruction into the network structure. Extensive experiments on the challenge MVTecLOCO, MVTecAD, VisA and MADsim datasets verify the advantage of proposed LogicAL on both logical and structural anomaly localization.

Figures (14)

• Figure 1: LogicAL is a novel anomaly synthesis framework for unsupervised anomaly localization. (a) and (b) are real normal and anomaly examples from MVTecLOCO mvtecloco dataset. As shown in (f), LogicAL can generate photo-realistic anomalies that are flawless in appearance but violate the underline logic constraints, such as adding extra component and causing mismatch among components. On the contrary, the synthetic anomalies using (c) cutpaste, (d) draem and (e) omnial are prone to be like structural anomalies that have faulty appearance.
• Figure 2: Framework of LogicAL. LogicAL converts a normal image into an anomaly image based on edge manipulation. (c) to (g) show that it generates logical anomaly image by removing normal edges from the selected semantic regions. It syntheses structural anomaly image by replacing normal edges in arbitrary regions with the augmented edges that are sampled from trainset, as shown in (h) to (k).
• Figure 3: Visualization of our synthetic anomalies. (a) and (b) are extracted from (f). (c) is generated by merging the normal edges (b) from the selected regions of (a) with sampled edges. Anomaly image (e) is converted from (c). (d) is the difference map between (e) and (f).
• Figure 4: Framework of proposed unsupervised anomaly localization. It consists of anomaly synthesis and localization modules. Anomaly synthesis is based on anomaly edge map construction and edge-to-image generation. Synthetic anomaly is reconstructed into normal image, corresponding just noticeable distortion (JND) and edge maps. Anomaly localization is achieved by exploring difference between reconstructed and original data.
• Figure 5: Edge-to-image generation. (a) Normal images from MADsim pad, MVTecAD mvtec and VisA visa, (b) Edge maps extracted by PiDiNet pidinet, (c) TPS tps warping on (b), (d) Images generated from (c) by DeepSIM deepsim, (e) and (f) are synthesis anomaly edge maps and images applied same TPS warping.