What it does This is a C++ computer-vision project that detects and recognizes Chinese license plates. It chains together classic and deep-learning techniques: Haar cascades or MTCNN for plate detection, perspective correction via corner-point regression, then character segmentation and recognition through CNNs, BLSTM+CTC, or FCN. The author claims 99.8% accuracy and ~100 ms per plate.

The interesting bit The project is essentially a survey course in plate recognition crammed into one repo. Rather than picking one approach, it implements three competing pipelines side by side—Haar+CNN, MTCNN+LSTM, and FCN—so you can trade off speed, accuracy, and data hunger. The MTCNN branch does perspective transform from detected corners, which the author notes handles badly angled plates.

Key highlights

• Three detection options: Haar cascade, MTCNN, or (planned) YOLO
• Three recognition backends: per-character CNN, sequence LSTM+CTC, or FCN with deconvolution
• Includes a 37 GB dataset (19 files, ~2 GB each) distributed via WeChat QR code
• Windows-only build: VS2015, x64, CUDA 9.2
• Third-party dependencies hosted on Baidu Pan, not in repo

Caveats

• Build instructions are Windows-centric and pinned to CUDA 9.2 (released 2018)
• YOLO and “jump-point” image-processing pipelines are listed but marked unsupported
• README is mostly a math tutorial (linear algebra through GANs) and WeChat marketing; actual code documentation is thin
• Dependency download requires Chinese cloud storage and a QQ group

Verdict Worth a look if you’re specifically targeting Chinese plates and need a reference implementation of multiple classical-to-DL pipelines. Skip it if you want something modern, cross-platform, or maintainable—this is a 2018 time capsule with a very specific toolchain.