12.3K ·GitHub
What is this?
An ultra-lightweight Chinese OCR engine: it finds and recognizes text in images, including vertical text. Detection, recognition and angle classification run on three tiny models that total only 4.7MB, entirely on CPU.
An ultra-lightweight Chinese OCR engine: it finds and recognizes text in images, including vertical text. Detection, recognition and angle classification run on three tiny models that total only 4.7MB, entirely on CPU.So what am I actually installing — a Python tool, a web app, or a mobile SDK?The main branch is Python: a CLI plus a Tornado web service with a browser page. The repo also carries independent reference demos for C++, JVM (Java/Kotlin), Android and .NET.
The main branch is Python: a CLI plus a Tornado web service with a browser page. The repo also carries independent reference demos for C++, JVM (Java/Kotlin), Android and .NET.How do I get it running?Install it as a Python package — the .onnx models ship with it, so the CLI works out of the box. Or start the Tornado web service and open its address in a browser; it listens on port 8089 by default.
Install it as a Python package — the .onnx models ship with it, so the CLI works out of the box. Or start the Tornado web service and open its address in a browser; it listens on port 8089 by default.What does my machine need to run it — a GPU or just a normal laptop?Python 3.6 on Windows, Linux or macOS. The main branch infers on CPU, so no CUDA or GPU setup is needed; GPU builds exist only in the NCNN demo variants.
Python 3.6 on Windows, Linux or macOS. The main branch infers on CPU, so no CUDA or GPU setup is needed; GPU builds exist only in the NCNN demo variants.Is Chinese the only language it can read?It is built for Chinese text — the README claims no other languages. Also, the C++, JVM, Android and .NET demos are reference ports translated from the Python version, each maintained independently rather than as polished products.
It is built for Chinese text — the README claims no other languages. Also, the C++, JVM, Android and .NET demos are reference ports translated from the Python version, each maintained independently rather than as polished products.Will it run on my phone?Android is covered by official demos: ONNX Runtime, NCNN (CPU and GPU) and MNN. iOS has no official demo — the third-party TNN demo adds iOS using ARM CPU, OpenCL or Metal backends.
Android is covered by official demos: ONNX Runtime, NCNN (CPU and GPU) and MNN. iOS has no official demo — the third-party TNN demo adds iOS using ARM CPU, OpenCL or Metal backends.If I feed it images, does everything stay on my machine?Yes — inference happens entirely on your machine's CPU with the bundled models; no external service is involved.
Yes — inference happens entirely on your machine's CPU with the bundled models; no external service is involved.Can it handle vertical Chinese text, like columns in traditional layouts?Yes — vertical text recognition is called out as a core feature in the project's own description, alongside the ultra-small 4.7MB total model size.
Yes — vertical text recognition is called out as a core feature in the project's own description, alongside the ultra-small 4.7MB total model size.I want the Android or C++ demo without wiring native deps myself — possible?Yes — the README says complete source projects are available in a QQ group's shared files if you don't want to integrate ONNX Runtime, NCNN or MNN yourself.
Yes — the README says complete source projects are available in a QQ group's shared files if you don't want to integrate ONNX Runtime, NCNN or MNN yourself.Can it give me the position of each individual character?There's a third-party project, fanqie03/char-detection, built on this repo and CRNN principles, that derives the position of each character and combines characters into words.
There's a third-party project, fanqie03/char-detection, built on this repo and CRNN principles, that derives the position of each character and combines characters into words.What are similar projects?ncnn, HivisionIDPhotos, manga-image-translator, YOLOX, X-AnyLabeling, openvino, doctr, SimdPaddleOCR
ncnn, HivisionIDPhotos, manga-image-translator, YOLOX, X-AnyLabeling, openvino, doctr, SimdPaddleOCRWhich topics is it in?ncnn, onnx, lightweight, ocr, inference, computer-vision, c++, pytorch, deep-learning, android, machine-learning, chinese
ncnn, onnx, lightweight, ocr, inference, computer-vision, c++, pytorch, deep-learning, android, machine-learning, chinese