PCB Inspection · Vision AI · Publication System

PCB Visual Inspection Studio

Quality checks, ROI registration, and anomaly detection for F411 PCB video, with PatchCore, FR-PatchCore, and DINOv3 feature studies.

Paper-style figures and homepage integrated

↗ Project Page

Open the files

PCB Visual Inspection

ROI registration · PatchCore, DINOv3

Methods and figures Result figures

Problem

The project turns camera intake, ROI alignment, anomaly heatmaps, masks, tables, paper figures, and homepage publishing into one coherent PCB inspection publication structure.

Architecture

USB/industrial camera or file intake → quality gate → ROI alignment → ViT/DINO/PatchCore-style features → anomaly score/heatmap/mask → reviewer gate → paper, GitHub, Notion, and homepage publishing.

Hardware and Software

Python
OpenCV
PyTorch
CUDA
PatchCore
FR-PatchCore
DINOv3
Jekyll
GitHub Pages
Notion

Data Flow

Raw frames and ROI evidence are kept separate, and each figure is reused as an independent square panel rather than a bundled contact sheet.

Methodology

Extracted layout grammar from the supplied figure bundles using a source-checked reference-intake route
Redrew text boxes, axis labels, arrows, color bars, and centered panels as square publication figures
Configured the homepage gallery around individually clickable enlarged images

Results

Separated ten PCB methodology figures into independent cards
Mirrored the local 8765 portal into a public GitHub Pages route
Verified click-to-enlarge behavior for the first methodology figure in the browser
Published the v9 gallery cache-bust and separated figure references in the PCB repo and public hub

Architecture and result views

PCB camera, video, ROI, and quality-gate flow drawn as a square paper figure — fig31a. capture and quality gate

PCB ViT, DINOv3, and patch-feature path drawn as a square paper figure — fig31b. ViT/DINO patch path

PCB registration and memory-bank flow drawn as a square paper figure — fig31c. registration and memory bank

PCB heatmap and mask fusion result panels drawn as a square paper figure — fig31d. heatmap and mask fusion

PCB source and synthetic domain boundary evidence drawn as a square paper figure — fig31e. domain boundary

PCB GitHub, Notion, and homepage publication loop drawn as a square paper figure — fig31f. publication hub loop

PCB anomaly-score distribution plot drawn as a square paper figure — fig31g. score distribution

PCB feature-embedding cluster plot drawn as a square paper figure — fig31h. embedding clusters

PCB loss-convergence plot drawn as a square paper figure — fig31i. loss convergence

PCB dataset-balance chart drawn as a square paper figure — fig31j. dataset balance

What I worked on

Removed non-PCB-inspection VIScan material
Reorganized camera/video intake, quality gate, ROI, feature extraction, review, and publication flow
Built a reusable single-figure atlas for paper, GitHub, Notion, and homepage reuse

Code and results

v9 separated gallery and click-to-enlarge behavior in the PCB public homepage
Ten 900×900 paper-style square figures, fig31a through fig31j
Full Research Publication Workbench mirrored into GitHub Pages
PatchCore, FR-PatchCore, DINOv3, and source-checked figure-redraw record

Design scope and next steps

Available records: research architecture, figures, and local experiments. Production-evaluation criteria: gold-label comparison, defect decisions, and yield.

Project resources

Project Pages

The F411 PCB inspection study connects frame-quality checks, ROI registration, feature extraction, and anomaly detection. The records include PatchCore, FR-PatchCore, and DINOv3 feature studies with method and result figures.