Inkive App Exports Paper-Book Highlights to Obsidian via On-Device OCR

The developer, whose name was not provided, said the app performs three steps: locating markings such as underlines, highlights, and circles; reading the text near those markings; and determining which sentence the reader intended to preserve. To find markings, the developer trained a Core ML image segmentation model using photos shared by reading bloggers, with varied paper textures, lighting, and shooting angles. The developer had no algorithm background and used AI tools to help prepare data, label samples, and adjust training. The model reached an accuracy of 0.94 and a recall of 0.88 on the test set, according to the post.

For text recognition, the developer initially relied on Apple's Vision framework but encountered problems with curved pages and uneven lighting. After seeing the release of PPOCR v6 in June, the developer switched to that model, finding it more stable for Chinese paper books. All processing is done on-device, meaning photos and results stay on the phone and no cloud fees are incurred.

The hardest part, the developer wrote, was deciding what the reader actually meant to highlight. Because handwritten markings are imprecise, the app does not need pixel-perfect alignment. Instead, the developer measured an "acceptability rate": whether the result matches the reader's judgment. In tests, 95% of results required no changes or only minor proofreading, a level the developer said made the app suitable for daily use.

The developer said they tried existing tools, including Highlighted, Baimiao, vFlat Scan, and ShelfMe, but none solved the problem. They also experimented with large language models such as Doubao, DeepSeek, ChatGPT, and Gemini, but results were inconsistent, sometimes missing the marked lines or including unmarked text. Inkive was created as an alternative that does not require changing reading habits: users read paper books as usual, then photograph the pages after finishing. The app lets users proofread each excerpt or preview the entire book before exporting a Markdown file.

The post also mentioned a hardware product called Mark II, made by an overseas startup, which uses a smart bookmark device to scan highlighted lines and syncs via cloud AI to Notion and Readwise. The developer said Inkive differs by using only a phone and local models, avoiding the need for dedicated hardware.

Inkive is now available on the App Store. Android, Web, and mini-program versions are planned, according to the developer.