JAEN

Studio Toriumi

Three ways to turn a book into text.

Digitising books you own is called jisui in Japan, and the law draws an unusual line: doing it yourself is legal, paying someone else to do it is not. This page covers that line first, then compares the three methods on time, cost and accuracy — and ends by breaking down what the "film yourself flipping pages" approach actually demands technically.

Read the guide JA

What's insideWhat's in it

The legal line, in three rules

Copying a book you own, for yourself, is permitted under Article 30 of Japan's Copyright Act. Paying a service to scan it for you was ruled infringing by the IP High Court in 2014 — the copying is done by the company, so it isn't private use. And sharing the resulting file, in any form, is out.

Cut and scan — fastest, best quality, destroys the book

Guillotine the spine, feed the loose pages through a document scanner. Five to ten minutes a book, and the quality beats everything else because flat paper needs no dewarping. Worth the equipment above roughly 30 books.

Shoot page by page — free, book survives, extremely slow

vFlat, Adobe Scan and the built-in iPhone scanner all split spreads, remove fingers and correct curvature well now. The problem is volume: a 300-page book means 150 shutter presses, and 30 to 60 minutes.

Film the flip — fast to capture, hard to process

Recording a page-flip takes one to two minutes per book, twenty to thirty times faster than shooting each page. It hasn't become the norm because the difficulty moved to post-processing, not because the idea is wrong.

Why video is hard, in five parts

Frame selection (a two-minute clip at 30fps is 3,600 frames, and only the still, fully-open ones are usable — Laplacian variance for blur detection), dewarping (the hardest part, and where products win or lose), deduplication and page ordering, capture conditions (lighting and shutter speed cannot be recovered in software), and OCR — which is no longer the bottleneck.

The bottleneck moved

Apple's Vision framework and Google's ML Kit now run OCR on-device, free, including vertical Japanese. You no longer build an OCR engine. The competition is entirely in the steps before it — which is a much more accessible problem for a small developer.

Practical accuracy tips

300dpi minimum. Two lights at an angle, never one from the front — direct light blows out the paper. Greyscale often beats binarisation on modern OCR. And specify the language and writing direction explicitly for vertical Japanese.

LanguageReading this in English

The full comparison, the method-selection calculator and the equipment list are written in Japanese. The calculator works numerically regardless of language.

How this site handles translation

We don't machine-translate pages and publish them as if they were written in English. When an English edition exists, it has been written in English. Until then, this page tells you honestly what is on the other side of the link.

Everything on this site is free and has no paywall, in either language.