Repository navigation
v0.3.7 — bounded curvature recovery
Recover additional curved QR codes by allowing one final strict refinement after all earlier recovery paths fail. The candidate must already be retained with exact version evidence and must not have been attempted at full size. Previous successful results and metadata keep priority.
| Accuracy set | 0.3.6 | 0.3.7 | ZXing-WASM 3.1.7 |
|---|---|---|---|
| Evaluated distortions | 37/96 | 44/96 | 42/96 |
| Earlier distortions | 81/128 | 89/128 | 106/128 |
| Separate set frozen before evaluation | 48/128 | 53/128 | 52/128 |
| Earlier curved sheets | 3/18 | 4/18 | Not run here |
| Transformed photographs | 2,872/2,944 | 2,872/2,944 | Not run here |
No previous read or metadata is lost and no wrong QR payload appears in these checks. The separate set is now regression data. ZXing remains substantially ahead on the earlier 128 distortions; these results do not establish a universal accuracy lead.
TypeScript, Rust, Kotlin and packaged WASM agree. Three new shared fixtures cover normal, inverted and mirrored fifth-attempt recovery. The maximum rises explicitly to five refinement attempts and 7,187,500 alignment probes; the original four-attempt order and map cap remain unchanged.
Two paired M4 runs keep main-noise WASM medians within 0.5% and clean within 0.4%. Four newly recovered curved examples cost 13.4–21.2% more than their earlier misses; the tested persistent curved miss costs 19.7–21.0% more. This release trades additional work on difficult frames for measured accuracy.
See the performance report, source-bound evidence and reproduction commands and installation guide.
Publication is verified across npm, JSR, crates.io, Maven Central, the documentation site and both CDNs. The release verification receipt links all seven successful workflows and six public readbacks. npm verification passed on its second attempt after the public archive became available; the package was uploaded once.