1. Card Database
Sync once for much better matching. The scanner uses this local index before trying online fallback search.
The first sync can take a while because it downloads card metadata into your browser. After that, scans are matched locally and exported from this device.
2. Accuracy Test
After syncing the database, run this on 100 known official card records/images. It checks whether the scanner returns the correct card ID, name, set, and number.
This test uses the same OCR and matching code as a real scan. It is slower, but it exposes whether the scanner is failing because of OCR, number parsing, database matching, visual matching, or card detection. The report includes the expected card data for every test card.
3. Scan One Card
Place the card inside the yellow guide. The scanner crops from that exact guide first, then runs card detection if needed.
4. Review Match
Accept only when the card image, name, set, and number look right.
Normalized card preview appears here after scanning.
5. CSV Export
Accepted scans are saved in this browser until you clear them.
| Name | Set | Number | Finish | Confidence | Review |
|---|---|---|---|---|---|
| No accepted scans yet. | |||||
What changed in V4
- Database-first matching: downloads real card/set metadata and builds number/set indexes locally.
- Number-first recognition: collector number like
013/198,TG01/TG30, and promo-style numbers are handled better. - Multiple OCR passes: top name, top-wide, bottom, bottom-left, bottom-right, and old symbol zone are read separately with different image cleanup methods.
- Candidate scoring: exact number + printed total + fuzzy name + set code + optional visual image check.
- Review-first safety: if the top match is not clearly better than the second match, it marks the scan for review.
- Built-in 50-card test: after syncing, it can test itself against 50 official card images and show exactly where recognition fails.
Browser-only scanning still has limits. For near-app-store-level accuracy, the next jump would be a backend with cached images, a trained set-symbol model, and stronger OCR/image matching.