opencv-kmp

Kotlin Multiplatform bindings for OpenCV (5.x, Android SDK API surface: core / imgproc / imgcodecs / video / videoio / objdetect / dnn / photo / calib / features2d / stereo / ptcloud).

• Kotlin/Native — OpenCV is compiled statically per target and embedded into the published klib through cinterop; consumers link it without any extra setup.
• JVM (desktop + Android) — a self-contained JNI shared library (libopencv_jni) is built from the OpenCV submodule and shipped as a classpath resource per OS/arch; NativeLoader extracts and loads the matching one at runtime.

Both platforms call the same thin C ABI (native/, cvk_ prefix) that wraps the C++ API, so behavior is identical everywhere. All native calls are exception-safe: failures surface as OpenCVException (with cv::Exception::what() text) or nullable returns.

import cn.enaium.opencv.*

fun main() {
    println(opencvVersion)

    mat(rows = 2, cols = 2, type = MatType.CV_32FC1, fill = Scalar.all(1.5)).use { a ->
        eye(2, 2, MatType.CV_32FC1).use { identity ->
            (a + identity).use { sum -> println(sum[0, 0]) } // 2.5
        }
        (a * 3.0).use { println(it[0, 0]) }                  // 4.5
    }

    imread("input.png")?.use { image ->
        val gray = image.cvtColor(ColorConversionCodes.BGR2GRAY)
        val edges = gray.canny(threshold1 = 50.0, threshold2 = 150.0)
        imwrite("edges.png", edges)
        gray.close()
        edges.close()
    }
}

Gradle:

dependencies {
    implementation("cn.enaium.opencv:opencv-kmp:<version>")
}

The JVM artifact bundles all nine JNI libraries (~9 × a few MB); Android apps get the right one automatically.

• Operators on Mat: + - * / with another Mat or a Scalar, indexed get/set, times(scale), broadcast + - / Double, abs()/squared(), infix bitwiseAnd/Or/Xor, diff, rsub.
• Scalar arithmetic: Scalar + - * / Scalar, * / + - Double, unaryMinus, infix dist.
• Extensions in common code: toGray(), toFloat32(), normalize01(), mirror(), rotate90/180/270(), pixels: ByteArray, shape, fill { }.
• Factories: mat(), zeros(), ones(), eye(), imread(), imwrite(), imencode(), imdecode().

Requirements: JDK 21+, CMake ≥3.16, a C++17 compiler; the OpenCV sources are a git submodule (git submodule update --init --recursive). Cross builds additionally need:

• linuxArm64 klib / linux JNI: aarch64-linux-gnu-gcc
• mingwX64 klib: x86_64-w64-mingw32-gcc
• androidNative klibs / android JNI libs: Android NDK (sdk.dir in local.properties, $ANDROID_HOME, or ~/Android/Sdk)

macOS klibs and the darwin JNI artifacts are built on macOS hosts; the windows-x86_64 JNI DLL is built on Windows hosts (MinGW via choco works).

# build + test everything buildable on this host
./gradlew :opencv-kmp:jvmTest :opencv-kmp:linuxX64Test   # platform tests
./gradlew :examples:basic:jvmRun                          # JVM demo
./gradlew :examples:basic:runDebugExecutableLinuxX64     # native demo

# publish local-platform artifacts to mavenLocal
./gradlew :opencv-kmp:publishToMavenLocal :jni-jvm-linux-x86_64:publishToMavenLocal

Apple mobile targets (iOS/tvOS/watchOS, simulator + device) compile but are not runnable from the CLI; the simulator suites run under iosSimulatorArm64Test / tvosSimulatorArm64Test / watchosSimulatorArm64Test (they need a booted simulator). The linuxArm64 klib is cross-compiled from x64 hosts, but Kotlin/Native has no linux-aarch64 host distribution in this Kotlin version, so there is no host that can run the linuxArm64 tests; the arm64 Linux JNI artifact is covered by the CI publish job instead.

Each Kotlin/Native target drives its own CMake configuration (configureNative_<target> / buildNative_<target>) that compiles OpenCV statically (BUILD_LIST=core,imgproc,imgcodecs,video,videoio,objdetect,dnn,photo,calib plus cvv, ptcloud-related modules, bundled zlib/libpng/libjpeg-turbo, no network downloads) plus the cvk shim, merges all archives into one libopencv_kmp.a, and embeds it into the klib.

OpenCV submodule patches. The vendored OpenCV tree ships with two patches applied automatically by native/CMakeLists.txt on every configure (idempotent, git apply --reverse --check first): native/cmake/mlas-subproject.patch (3rdparty/mlas resolve paths under add_subdirectory and Apple architecture detection from CMAKE_OSX_ARCHITECTURES) and native/cmake/imgcodecs-appletv.patch (exclude macOS-only imgcodecs sources on tvOS/watchOS). Keep the submodule checkout pristine; the build applies what it needs.

Two manually-dispatched workflows (Actions tab):

• Test — builds/tests every platform on matching runners and publishes to Maven Local.
• Publish — releases every artifact to Maven Central. The version is fixed in build.gradle.kts; nothing to input.

Required secrets for Publish: MAVEN_CENTRAL_USERNAME, MAVEN_CENTRAL_PASSWORD, SIGNING_KEY (base64 .gpg), SIGNING_KEY_ID, SIGNING_PASSWORD.

MIT — see LICENSE. OpenCV itself stays under its Apache 2.0 license.