AnomalyPCB：面向 PCB 邏輯與結構異常檢測的資料集建立

Abstract

工業異常定位（Industrial Anomaly Localization）由於像素級標註稀缺，在高解析度且元件密集分佈的印刷電路板組裝（PCBA）圖像中仍面臨挑戰。本文提出一個源自真實 SMT 生產線的大規模數據集 AnomalyPCB，其特點在於包含空間對齊的元件圖像，涵蓋邏輯、結構和元件級異常的多樣化缺陷。為確保質量同時擴展標註規模，我們提出了一個自動標註工作流程，包含偽標籤生成（Pseudo-label Generation）、人工質量檢驗與迭代重訓練機制。此外，我們提出了一個用於自動掩碼生成的基礎架構, 將參數高效微調（Parameter-Efficient Fine-Tuning）後的 SAM2 與於空間對齊的異常定位模塊相結合。實驗顯示，現有的少樣本方法（Few-shot Methods）難以應對 PCBA 圖像中特定的領域複雜性。因此 AnomalyPCB 可作為工業檢測研究中其一具挑戰性的基準。

Industrial anomaly localization remains a significant challenge due to the scarcity of pixel-level annotations, particularly in high-resolution, densely populated printed circuit board assembly (PCBA) imagery. In this work, we introduce AnomalyPCB, a large-scale dataset curated from online SMT production lines. It is characterized by spatially aligned components and heterogeneous defect patterns, including logical, structural, and component-level anomalies. To scale annotation while maintaining quality, we propose an iterative labeling pipeline that integrates pseudo-label generation, quality refinement, and iterative retraining.Furthermore, we establish a baseline framework for automatic mask generation by leveraging a parameter-efficiently adapted SAM2 coupled with a spatially aligned anomaly localization module. Extensive experiments reveal that existing few-shot methods struggle with the domain-specific complexities of PCBA images. Consequently, AnomalyPCB establishes a novel, challenging benchmark for advancing future research in industrial inspection.