Few Shot Part Segmentation Reveals Compositional Logic for Industrial Anomaly Detection

TL;DR

The paper tackles industrial anomaly detection with a focus on logical anomalies that require precise semantic segmentation of components. It introduces PSAD, which couples a few-shot, coordinate-aware segmentation model with three memory banks—class histogram, component composition, and patch representations—to detect both logical and structural anomalies. An adaptive scaling scheme unifies heterogeneous anomaly scores into a single robust metric, achieving state-of-the-art AUROC on the MVTec LOCO AD benchmark, notably 98.1% for logical anomalies. The approach demonstrates strong performance with limited labeled data and offers practical benefits for real-world manufacturing settings by leveraging prior knowledge of component composition and arrangement.

Abstract

Logical anomalies (LA) refer to data violating underlying logical constraints e.g., the quantity, arrangement, or composition of components within an image. Detecting accurately such anomalies requires models to reason about various component types through segmentation. However, curation of pixel-level annotations for semantic segmentation is both time-consuming and expensive. Although there are some prior few-shot or unsupervised co-part segmentation algorithms, they often fail on images with industrial object. These images have components with similar textures and shapes, and a precise differentiation proves challenging. In this study, we introduce a novel component segmentation model for LA detection that leverages a few labeled samples and unlabeled images sharing logical constraints. To ensure consistent segmentation across unlabeled images, we employ a histogram matching loss in conjunction with an entropy loss. As segmentation predictions play a crucial role, we propose to enhance both local and global sample validity detection by capturing key aspects from visual semantics via three memory banks: class histograms, component composition embeddings and patch-level representations. For effective LA detection, we propose an adaptive scaling strategy to standardize anomaly scores from different memory banks in inference. Extensive experiments on the public benchmark MVTec LOCO AD reveal our method achieves 98.1% AUROC in LA detection vs. 89.6% from competing methods.

Figures (5)

• Figure 1: Comparison of approaches at a conceptual level. (A) The anomaly detection (AD) model is directly trained using images. (B) Our proposed method guides part segmentation models using a few labeled samples to accurately segment components and then uses the segments for AD. (C) Examples of logical anomalies show the importance of semantically segmenting components for detection.
• Figure 2: Illustration of PSAD (Part Segmentation-based Anomaly Detection). During training depicted in the blue box, 3 different memory banks are constructed using normal images. The anomaly score of a test image is computed by finding its nearest neighbor (NN search) and adaptive scaling.
• Figure 3: Proposed part segmentation model that predicts segmentation utilizing visual and positional features.
• Figure 4: Qualitative comparison of FSS models. $\mathcal{L}_{sup}$ ($=\mathcal{L}_{CE}+\mathcal{L}_{Dice}$) denotes a supervised loss for labeled images. For the unsupervised methods, such as SCOPS and Part-Assembly, we arbitrarily set the number of parts as 10.
• Figure 5: Histogram visualizations of anomaly scores from different memory banks and the unified anomaly scores.