OpenCV.jl

OpenCV.jl is a Julia interface to the OpenCV computer-vision library — image and video processing, object detection, feature extraction, calibration, and more — backed by the precompiled library through OpenCV_jll.

📖 Documentation — a quick-start example, the Mat image type and its (channels, cols, rows) layout, the cv:: → OpenCV. naming convention, reading/writing images, and the API reference.

Installation

] add OpenCV

This installs the package and the precompiled OpenCV binaries; no system OpenCV is required.

CUDA

On Linux with an NVIDIA GPU and a CUDA 13 driver, OpenCV can use binaries built with CUDA and cuDNN, which enable the DNN_BACKEND_CUDA backend for cv.dnn. Install them next to OpenCV and restart Julia:

] add OpenCV_CUDA_jll

OpenCV then loads the CUDA build whenever it has binaries for the system, and falls back to the CPU build otherwise. Remove OpenCV_CUDA_jll to go back to the CPU build. It is a separate package because installing it downloads the CUDA runtime and cuDNN (about 2 GB).

OpenCV finds OpenCV_CUDA_jll in any environment on the load path, so installing it in the default environment (@v1.x) enables the CUDA build for every project. OpenCV ignores it, with a warning, if its OpenCV version differs from that of OpenCV_jll.

Developing locally

To hack on OpenCV.jl, check out the repository and dev it from Julia:

pkg> dev /path/to/OpenCV.jl   # or: pkg> dev OpenCV
pkg> test OpenCV

using OpenCV pulls in a matching OpenCV_jll automatically. Changes under src/ are picked up on the next using (or immediately with Revise). Changes under gen/ take effect only after regenerating the wrappers; rebuilding OpenCV_jll is needed only when the underlying OpenCV version changes.

Contributing

Contributions are welcome — please open issues and PRs. The codebase has three layers:

• Hand-written Julia in src/ — the module entry (OpenCV.jl), the Mat/Vec array types, the cpp↔julia conversions (mat_conversion.jl, types_conversion.jl), the geometry structs (typestructs.jl), errors.jl, show.jl, fileio.jl, and the videoio.jl conveniences. Edit these for the Julia-side API and array interop.
• Auto-generated wrappers in src/generated/ — produced by the generator under gen/. Do not hand-edit these. Regenerate with julia gen/regenerate.jl; see gen/PROVENANCE.md for the determinism guarantees and the OpenCV-bump checklist. cv_cxx_wrap.jl ends with a hand-authored "Manual Wrapping" section (GUI callbacks, CascadeClassifier) that the header parser cannot derive; its source is gen/binding_templates_jl/manual_cxx.jl, paired with the C++ side in gen/binding_templates_cpp/cv_core.cpp.
• The binding generator in gen/ — vendored from opencv_contrib/modules/julia with JuliaImages patches folded in. Edits here change src/generated/ on the next regeneration.

The compiled libopencv_julia lives in OpenCV_jll, built from Yggdrasil's O/OpenCV/build_tarballs.jl; it must come from the same OpenCV tag pinned in gen/OPENCV_VERSION, or @wrapmodule fails on mismatched symbols.

License

MIT — see LICENSE.

Acknowledgments

OpenCV.jl is built on the work of the OpenCV community. @archit120 was instrumental in developing this package during GSoC 2020 and 2021.