ldas.jp
JA
All tags

TAG

#onnx

6articles

📜Porting Kuzushiji Character Detection and Classical-Japanese OCR to JS/Next.js Without Python — ONNX Quantization and Library Extraction

A record of porting kuzushiji/historical-document character detection (YOLOv11) and NDL's koten-OCR-Lite (RTMDet + PARSeq) to JS/Next.js with onnxruntime-node — no Python. Covers ONNX export, SAHI, line detection, an int8 quantization accuracy comparison, why it is hard to run on Vercel, and extracting the core into a shared library.

kuzushijiocronnxonnxruntime

🧠Diagnosing and Fixing Memory-Related Crashes in an iOS OCR App

Investigation and fixes for a 6.7% crash rate in KotenOCR, an iOS OCR app using ONNX Runtime with 6 models totaling ~230MB on disk.

iosswiftmemoryonnx

🔍KotenOCR: Fixing Duplicate Detections in Modern-Script OCR Mode (Adding NMS and Class Filtering)

Fixing duplicate detection results in KotenOCR's modern-script OCR mode (NDLOCR-Lite) by adding NMS and filtering to line_* classes, achieving detection accuracy on par with the reference ndlocr-lite implementation.

iosocrndlonnx

📜KotenOCR v1.3.0: Dual OCR Modes for Classical and Modern Japanese Text

KotenOCR v1.3.0 adds NDLOCR-Lite support for modern printed text OCR alongside the existing classical text (kuzushiji) mode, with one-tap mode switching and parallelized recognition for up to 6.7x speedup.

iosocrkuzushijindl

⚡Parallelizing OCR Recognition on iOS with Swift Concurrency for up to 6.7x Speedup

Using withThrowingTaskGroup to parallelize OCR recognition in an iOS app, achieving 1.4x to 6.7x speedup depending on the number of detected text regions.

iosswiftconcurrencyocr

📜KotenOCR: An Offline iOS App for Recognizing Classical Japanese Cursive Script

KotenOCR is a free iOS app that runs the NDL Koten OCR-Lite model entirely on-device, enabling offline recognition of kuzushiji (classical Japanese cursive script) from photos.

iosocrkuzushijiswift