OCR engines need text at least 8-12 px tall, and upscaling small screenshots only partly helps

A test of optical character recognition on clean black-on-white Chinese and English text found that each engine has its own minimum text height, with no upper limit observed up to 60 px. Accuracy from the browser-based recognizer on Chinese was error-free from 12 px for Songti and 10 px for Heiti, while macOS needed 16 px for Songti and Tesseract needed 16 px for English. Near the floor, the font mattered almost as much as the engine, since thin serif horizontals shrank to about a pixel and caused more errors. Upscaling tiny images with bilinear interpolation recovered a lot for some engines, such as Tesseract on 8 px English, but could not restore strokes that had already merged at very small sizes. The author suspects that Lanczos interpolation may cause ringing artifacts, though this was not verified, and the reason an unscaled small image on a large canvas lowered one engine's accuracy remains unexplained.