Systematic Diagnosis of Cross-Domain Performance Degradation in PCB Defect Detection: A PKU-Source Study — Data, Analysis Outputs, and Reproducibility Resources
Description
This repository provides the data, analysis outputs, protocols, source data, and reproducibility resources supporting the study “Systematic Diagnosis of Cross-Domain Performance Degradation in PCB Defect Detection: A PKU-Source Study.” The study uses PKU as the sole source domain and MIXED-clean, SELF, and DeepPCB as observational target domains under a common five-class ontology consisting of mouse_bite, open_circuit, short, spur, and spurious_copper. The repository contains the study-specific dataset manifests and ontology mappings, experimental and evaluation protocols, annotation-audit and adjudication records, full-domain evaluation outputs, positive-control results, controlled-imaging and selected-target analyses, domain-characterization and robustness outputs, statistical results, and source data underlying the supplementary tables. It also provides the custom scripts, notebooks, configurations, environment specifications, and other resources required to trace and reproduce the reported analyses. Third-party public PCB datasets are not redistributed unless their original licenses explicitly permit redistribution. Their original sources are identified in the associated manuscript and repository documentation. Historical or superseded experimental branches are excluded from the manuscript-facing evidence package unless explicitly retained and labeled for provenance or audit purposes. The repository is intended to provide a traceable link from the manuscript claims to the corresponding protocols, code, formal results, supplementary source data, and provenance records.
Files
Steps to reproduce
Read the repository README, publication lock, artifact-status registry, and file manifest to identify the formal manuscript-facing artifacts and their provenance. Obtain the public source datasets from the original sources cited in the associated manuscript. Use the dataset manifests and native-to-canonical ontology mappings supplied in this repository to construct the formal common-five evaluation space. Reproduce the formal source-domain and cross-domain evaluations using the released model configurations, frozen evaluation settings, seeds, checkpoints or checkpoint identifiers, and environment specifications provided in the repository. Target-domain data must not be used for source-model selection or target-specific tuning. Run the released scripts for the evaluation-chain positive controls, annotation validation and frozen-prediction rescoring, controlled-imaging analysis, selected-target analysis, inference-sensitivity analysis, size analysis, and full-domain domain characterization according to the corresponding protocol files. Generate the manuscript tables, figures, and supplementary Tables S1–S4 from the released formal result files and source-data scripts. Do not substitute superseded or historical experimental branches for artifacts identified as current formal evidence. Verify the generated outputs against the repository manifests, formal locks, and SHA-256 records where provided. Detailed file-level dependencies, software versions, execution commands, and manuscript-to-resource mappings are provided in the repository README and associated protocol files.
Institutions
- Guizhou University of Engineering ScienceGuizhou, Bijie