UniPCB: A Unified Vision-Language Benchmark for Open-Ended PCB Quality Inspection

TL;DR

UniPCB tackles the lack of standardized evaluation for PCB quality inspection by creating a large-scale, multi-scenario vision-language benchmark and a domain-specific MLLM (PCB-GPT). The approach combines a systematic data construction pipeline with unified taxonomies and a three-stage curriculum training regime, including GRPO-based reinforcement learning, to achieve fine-grained defect localization and verifiable reasoning. Key findings show existing MLLMs struggle with PCB localization, while PCB-GPT significantly improves localization and structured outputs, achieving higher F1 and more robust QA across tasks. The work enables fair cross-domain comparison and provides resources to accelerate industrial PCB inspection research.

Abstract

Multimodal Large Language Models (MLLMs) show promise for general industrial quality inspection, but fall short in complex scenarios, such as Printed Circuit Board (PCB) inspection. PCB inspection poses unique challenges due to densely packed components, complex wiring structures, and subtle defect patterns that require specialized domain expertise. However, a high-quality, unified vision-language benchmark for quantitatively evaluating MLLMs across PCB inspection tasks remains absent, stemming not only from limited data availability but also from fragmented datasets and inconsistent standardization. To fill this gap, we propose UniPCB, the first unified vision-language benchmark for open-ended PCB quality inspection. UniPCB is built via a systematic pipeline that curates and standardizes data from disparate sources across three annotated scenarios. Furthermore, we introduce PCB-GPT, an MLLM trained on a new instruction dataset generated by this pipeline, utilizing a novel progressive curriculum that mimics the learning process of human experts. Evaluations on the UniPCB benchmark show that while existing MLLMs falter on domain-specific tasks, PCB-GPT establishes a new baseline. Notably, it more than doubles the performance on fine-grained defect localization compared to the strongest competitors, with significant advantages in localization and analysis. We will release the instruction data, benchmark, and model to facilitate future research.

Figures (6)

• Figure 1: Comparison between general industrial products and PCB inspection. (a) General products with obvious surface defects. (b) Unique PCB challenges: dense patterns, defect co-occurrence, and subtle cues. (c) Diverse imaging modalities are required to capture different defect characteristics.
• Figure 2: Overview of UniPCB Benchmark. We summarize the three annotation scenarios, unified defect and component taxonomies, task by target type, and the overall task proportions, with representative QA examples shown at the bottom.
• Figure 3: Statistical distribution of defect and component instances under the unified taxonomy.
• Figure 4: Data construction pipeline. We unify data sources, annotations, taxonomies, and a 14-type task schema, and construct two generation branches: a CoT-enabled training set and a multi-scenario benchmark. Each branch applies dual-track quality control with iterative prompt/rule refinement.
• Figure 5: Rings from inner to outer represent imaging modalities (RGB/AOI/Real), data construction stages, and task families with fine-grained sample counts.