ChineseOCR Lite: Ultra‑Lightweight OCR Engine for Vertical Chinese Text

Optical character recognition (OCR) technology is widely used, such as extracting key fields from invoices or recognizing questions in educational apps.

A new open‑source project called chineseocr_lite provides an ultra‑lightweight Chinese OCR engine that supports vertical text and runs on Linux/macOS using ncnn inference. The combined model size (enet + crnn + anglenet) is only 17 MB.

The project has already received over 2,400 stars on GitHub.

It builds on chineseocr and psenet to achieve Chinese natural‑scene text detection and recognition.

Key Features

Lightweight backbone detection model psenet (8.5 MB), crnn_lstm_lite (9.5 MB) and line‑text direction classification network (1.5 MB).

Arbitrary‑direction text detection with line‑direction classification during recognition.

crnn / crnn_lite LSTM/Dense recognition (ocr‑dense and ocr‑lstm are ports from chineseocr).

Support for vertical text recognition.

ncnn implementation of psenet (without kernel expansion).

ncnn implementation of crnn_dense (fully‑connected layers replaced by 1×1 convolutions).

ncnn implementation of shuuflenev2 angle classification network.

Complete OCR pipeline implemented with ncnn.

Recent Updates

Contributor nihui added crnn_lstm inference.

Upgraded model crnn_lite_lstm_dw.pth to crnn_lite_lstm_dw_v2.pth for higher accuracy.

Provided vertical‑text samples and a font library (fonts rotated 90°).

Font Style Example

Generated Vertical Text Sample

Recognition Effect

ncnn Detection & Recognition Demo (x86 CPU, single process)

Project repository: https://github.com/ouyanghuiyu/chineseocr_lite