Source-linked AI summary
PCBnet: A Dataset and Automatic Construction of SPICE Netlists from Schematic Images
Zhen Huang, Yuhao Gao, Yuzhi Liu, Daian Cheng, Chengyuan Shao, Yucheng Chen, Yongjian Jia, Futing Zhang, Yichen Shi, Wenhao Wang, Zuyan He, Yangbo Wei, Zhanfei Chen, Jinlong Yan, Yu Zhang, Haoying Wu, Ting-Jung Lin, Lei He
TL;DR
AI-driven PCB design automation lacks large-scale paired schematic–netlist data, despite the complexity of PCB schematics. PCBnet supplies a richly annotated dataset and combines visual recognition, topology construction, and multi-agent correction, achieving 84.47% connectivity accuracy alongside high component and text recognition scores.
Problem
PCBnet addresses the lack of large-scale schematic–netlist paired data for PCB circuit-topology modeling amid diverse components, complex wiring, and noisy annotations.
Method
The paper constructs PCBnet with annotated schematic images and paired SPICE netlists, then uses visual recognition, topology construction, and domain-knowledge-guided multi-agent correction.
Results
94.54% component detection mAP, 98.57% text recognition accuracy, and 84.47% connectivity accuracy are achieved on PCBnet.
Takeaways & Limitations
PCBnet establishes a benchmark and data foundation for future AI-driven PCB design automation.
Abstract
from arXiv · showhide
Printed circuit boards (PCBs) are fundamental to modern electronic systems, yet AI-driven PCB design automation remains constrained by the lack of large-scale paired schematic-netlist datasets. PCB schematics are particularly challenging due to diverse component types, complex wiring topologies, and noisy textual annotations. To address this gap, we present PCBnet, a large-scale PCB schematic dataset comprising over 300 real-world designs with annotated pins and paired SPICE netlists. It contains more than 50,000 component instances, 150,000 wires, 100,000 text regions, and 400,000 characters. We further develop an automated schematic-to-netlist pipeline that combines visual recognition, topology construction, and domain-knowledge-guided multi-agent correction. The proposed method achieves 94.54% component detection mAP, 98.57% text recognition accuracy, and 84.47% end-to-end connectivity accuracy. PCBnet provides a benchmark and data foundation for future AI-driven PCB design automation.
I. INTRODUCTION
PCBnet addresses the lack of large-scale paired schematic–netlist data for PCB design by introducing a richly annotated dataset and an automated schematic-to-netlist framework. The framework combines visual recognition, topology construction, and multi-agent correction, achieving strong component, text, and connectivity performance.
- PCB design remains largely manual because of diverse component types, complex wiring topologies, and heterogeneous design conventions.
- Existing PCB datasets largely target defect detection, leaving no large-scale schematic–netlist paired dataset for modeling circuit topology.
- PCBnet contains over 300 real-world PCB designs with annotations for components, pins, wires, texts, and paired SPICE netlists.
- 94.54% component detection mAP, 98.57% text recognition accuracy, and 84.47% connectivity accuracy are achieved on PCBnet.
- The schematic-to-netlist framework combines visual recognition, structure construction, and multi-agent correction to recover circuit connectivity from schematic images.
- PCBnet provides benchmarks for component detection, text recognition, and netlist construction, forming a data foundation for AI-driven PCB design research.
II. DATASET CONSTRUCTION
PCBnet is built from diverse real-world schematic files with one-to-one image–netlist correspondence and multi-level annotations. Each schematic is represented as a circuit graph whose nodes are component pins and whose edges are electrical connections inferred from pin–wire proximity.
- PCBnet uses over 300 real-world schematic files from open-source hardware projects and EDA tools including KiCad and EasyEDA.
- Each design pairs a rendered schematic image with the corresponding SPICE netlist extracted from the original engineering file.
- Schematic annotations include component bounding boxes, pin locations, text regions, and wire masks, with component and text labels manually corrected for accuracy.
- Each schematic is represented as a circuit graph G = (V, E), where V denotes component pins and E denotes electrical connections.
- Pin–wire associations are established using a distance threshold, then used to construct the circuit graph and generate the SPICE netlist.
III. IMAGE-TO-NETLIST PIPELINE
The image-to-netlist pipeline reconstructs circuit topology from a schematic image through visual element extraction, circuit structure construction, and multi-agent correction. Its extraction stage detects components and text while retaining OCR confidence for recognized labels.
- The framework consists of Visual Element Extraction, Circuit Structure Construction, and Multi-agent Correction.
- Visual Element Extraction: YOLOv11 detects components as bounding boxes with category labels.
- Visual Element Extraction: YOLOv11 localizes text regions and PaddleOCR recognizes their contents, testing a 90°-rotated crop for vertically oriented labels.
- Visual Element Extraction: Each recognized text item is represented by its token, bounding box, and OCR confidence score.
1) Wire Segmentation:
The framework extracts wires and components, infers pin-based circuit connectivity, and matches textual labels to components using spatial and schematic-specific constraints before netlist conversion.
- Wire Segmentation: Wire segmentation predicts a binary wire mask and skeletonizes it into one-pixel-wide wiring representations.U-Net produces M = Seg(I), where wire pixels are marked in the binary mask.
- Construction: Pin locations are estimated from component categories and their geometry, leveraging consistent configurations of common components.Resistors, capacitors, and integrated circuits provide typical layouts for geometry-guided candidate pin locations.
- Construction: Connectivity inference treats pins as graph nodes and links them to spatially adjacent wire segments, yielding G = (V, E).The graph edges represent inferred electrical connections.
- Text-Component Matching: Text-component matching begins with nearest-neighbor association and combines spatial proximity with naming conventions and layout characteristics.Nearest-neighbor matching is effective when component identifiers are placed close to their corresponding components.
- Text-Component Matching: Category consistency, geometric alignment, and local uniqueness constraints reduce ambiguous text assignments before topology and SPICE netlist construction.These constraints address cross-type mismatches, layout alignment, and conflicting local assignments.
- Pipeline Output: The pipeline organizes extracted components, pins, wires, and labels into a structured topology and converts it into a SPICE netlist.The overview comprises visual extraction, structure construction and matching, and multi-agent correction.
C. Stage 3: Multi-agent Correction
Stage 3 addresses structural inaccuracies caused by recognition errors by applying multi-agent correction supported by domain-specific knowledge.
- Error Motivation: Recognition errors, especially in text, can propagate from Stage 2 into misidentified components and incorrect net labels.These errors create structural inaccuracies in the initial circuit topology.
- Correction Framework: The correction framework targets OCR errors such as missing underscores and confusion among visually similar characters.It integrates recognition results with domain-specific knowledge because visual processing alone may not resolve these patterns.
1) Confidence-Guided Gatekeeper:
Stage 3 uses confidence to route uncertain OCR tokens through progressively richer correction, combining domain rules with visual-model candidates while limiting extra computation.
- Confidence-Guided Gatekeeper: Predictions with c_i ≥ τ are accepted directly, while low-confidence tokens are routed to correction.The gatekeeper focuses computation on uncertain cases and avoids unnecessary overhead.
- Rule-Based Correction: For uncertain tokens, an LLM correction agent retrieves naming conventions, character-confusion patterns, and representative examples from a domain knowledge base.The system retrieves relevant rules R_i for constrained generation.
- Rule-Based Correction: The constrained correction design restricts outputs to valid naming patterns and is described as improving correction accuracy.The corrected token is selected through a probability-based generation step conditioned on the OCR token and retrieved rules.
- Multi-modal Image Correction: When rule-based correction is insufficient, a vision-language model receives the cropped text image to produce a candidate.This multimodal step directly leverages visual information.
- Confidence Selection: The final prediction is selected by confidence, retaining the multimodal candidate when its confidence exceeds that of the original token.Most tokens are already correct, so this correction applies only to a small subset and adds limited overhead.
4) Component-Aware Pairing:
Component-aware pairing uses the matched component category as a semantic constraint when correcting low-confidence tokens, reducing ambiguity in naming predictions.
- Component-Aware Pairing: Each low-confidence token is paired with its associated component from text-component matching before correction.This pairing supplies component context to the correction process.
- Component-Aware Pairing: The corresponding component category constrains candidate token generation, such as R for resistors and C for capacitors.The category is denoted by l_j in the formulation.
- Outcome: Integrating visual extraction, structure construction, and multi-agent correction enables reliable and consistent schematic-to-netlist construction.Component-aware semantic constraints improve correction accuracy in ambiguous cases.
IV. EXPERIMENTS
The experiments evaluate the full pipeline on component detection, text recognition, and netlist construction, with component detection achieving strong performance across diverse schematic layouts. General-purpose multimodal models perform substantially worse on precise component localization.
- Evaluation Setup: The full pipeline is evaluated on component detection, text recognition, and netlist construction.Component detection uses mAP; text recognition uses CER, WER, and accuracy; netlist construction uses connectivity accuracy.
- Component Detection: 94.54% mAP is achieved for component detection across diverse PCB schematic layouts.The evaluation covers variation in component types, symbol styles, scales, orientations, and dense overlapping visual elements.
- Component Detection: General-purpose multimodal models achieve around 30% component accuracy even on cropped component regions.On full schematic images, these models often fail to reliably localize individual components in dense, fine-grained layouts.
C. Text Recognition Results
Text recognition improves substantially when text regions are localized before recognition, and the multi-agent correction modules further reduce recognition errors. YOLO+PaddleOCR with all correction modules achieves the strongest reported recognition results.
- Detection-Assisted OCR: Full-image OCR performs poorly, with CER typically above 20% and WER above 40%.Direct OCR is susceptible to background noise, dense layouts, and non-text elements.
- Detection-Assisted OCR: YOLO+PaddleOCR achieves 6.33% CER, 19.04% WER, and 93.67% ACC in the detection-assisted setting.The result supports separating text detection from recognition in complex schematic images.
- Multi-Agent Correction: Each correction module improves recognition performance, reducing CER from 6.33% to 1.43% and WER to 3.59%.The knowledge-guided LLM, multimodal image agent, and component-aware pairing reduce successive recognition errors.
- Multi-Agent Correction: 98.57% overall accuracy is achieved after adding the complete multi-agent correction mechanism.The multimodal image agent lowers CER to 2.52% before component-aware pairing produces the final result.
D. Netlist Construction Results
The end-to-end system constructs PCB circuit topology and generates structured netlists from schematic images. It achieves high wire extraction accuracy and 84.47% connectivity accuracy, while general-purpose multimodal models perform substantially worse.
- End-to-End Construction: 93.70% wire extraction accuracy and 84.47% connectivity accuracy are achieved for end-to-end netlist construction.Connectivity accuracy measures the proportion of correctly recovered pin-to-pin connections.
- Comparison: General-purpose multimodal models achieve approximately 30% component accuracy and below 10% connectivity accuracy.The comparison concerns their ability to capture the fine-grained geometric and topological structures of PCB schematics.
- Framework: The framework integrates visual recognition, structure construction, and multi-agent correction to recover circuit connectivity.The resulting topology-oriented pipeline supports schematic-to-netlist construction.