wangzhengyang 718c41634f feat: 切换后端至PaddleOCR-NCNN，切换工程为CMake

1.项目后端整体迁移至PaddleOCR-NCNN算法，已通过基本的兼容性测试
2.工程改为使用CMake组织，后续为了更好地兼容第三方库，不再提供QMake工程
3.重整权利声明文件，重整代码工程，确保最小化侵权风险

Log: 切换后端至PaddleOCR-NCNN，切换工程为CMake
Change-Id: I4d5d2c5d37505a4a24b389b1a4c5d12f17bfa38c

2022-05-10 10:22:11 +08:00

11 KiB

Raw Blame History

opencv-mobile

✔️ This project provides the minimal build of opencv library for the Android, iOS and ARM Linux platforms.

✔️ Packages for Windows, Linux, MacOS and WebAssembly are available now.

✔️ We provide prebuild binary packages for opencv 2.4.13.7, 3.4.16 and 4.5.4.

✔️ We also provide prebuild binary package for iOS with bitcode enabled, that the official package lacks.

✔️ All the binaries are compiled from source on github action, no virus, no backdoor, no secret code.

opencv 4.5.4 android	package size
The official opencv	235MB
opencv-mobile	16.1MB

opencv 4.5.4 ios	package size
The official opencv	183MB
opencv-mobile	15.4MB

opencv 4.5.4 ios with bitcode	package size
The official opencv	missing :(
opencv-mobile	53.7MB

Download

Android

(armeabi-v7a, arm64-v8a, x86, x86_64) build with ndk r21d and android api 24.

iOS

(armv7, arm64, arm64e, i386, x86_64) build with Xcode 12.4.

iOS with bitcode

(armv7, arm64, arm64e, i386, x86_64) build with Xcode 12.4.

ARM Linux

(arm-linux-gnueabi, arm-linux-gnueabihf, aarch64-linux-gnu) build with ubuntu cross compiler.

Windows

(x86, x64) build with VS2015, VS2017 and VS2019.

Linux

(x86_64) build on Ubuntu-18.04 and 20.04.

MacOS

(x86_64, arm64) build with Xcode 12.4.

WebAssembly

(basic, simd, threads, simd+threads) build with Emscripten 2.0.8.

Usage Android

Extract archive to <project dir>/app/src/main/jni/
Modify <project dir>/app/src/main/jni/CMakeListst.txt to find and link opencv

set(OpenCV_DIR ${CMAKE_SOURCE_DIR}/opencv-mobile-4.5.4-android/sdk/native/jni)
find_package(OpenCV REQUIRED)

target_link_libraries(your_jni_target ${OpenCV_LIBS})

Usage iOS and MacOS

Extract archive, and drag opencv2.framework into your project

Usage ARM Linux, Windows, Linux, WebAssembly

Extract archive to <project dir>/
Modify <project dir>/CMakeListst.txt to find and link opencv

set(OpenCV_DIR ${CMAKE_SOURCE_DIR}/opencv-mobile-4.5.4-armlinux/arm-linux-gnueabihf/lib/cmake/opencv4)
find_package(OpenCV REQUIRED)

target_link_libraries(your_target ${OpenCV_LIBS})

How-to-build your custom package

step 1. download opencv source

wget -q https://github.com/opencv/opencv/archive/4.5.4.zip -O opencv-4.5.4.zip
unzip -q opencv-4.5.4.zip
cd opencv-4.5.4

step 2. strip zlib dependency and use stb-based highgui implementation (optional)

patch -p1 -i ../opencv-4.5.4-no-zlib.patch
truncate -s 0 cmake/OpenCVFindLibsGrfmt.cmake
rm -rf modules/gapi
rm -rf modules/highgui
cp -r ../highgui modules/

step 3. patch opencv source for no-rtti build (optional)

patch -p1 -i ../opencv-4.5.4-no-rtti.patch

step 4. apply your opencv options to opencv4_cmake_options.txt

step 5. build your opencv package with cmake

mkdir -p build
cd build
cmake -DCMAKE_INSTALL_PREFIX=install \
  -DCMAKE_BUILD_TYPE=Release \
  `cat ../../opencv4_cmake_options.txt` \
  -DBUILD_opencv_world=OFF ..

step 6. make a package

zip -r -9 opencv-mobile-4.5.4.zip install

Some notes

The minimal opencv build contains most basic opencv operators and common image processing functions, with some handy additions like keypoint feature extraction and matching, image inpainting and opticalflow estimation.
Many computer vision algorithms that reside in dedicated modules are discarded, such as face detection etc. You could try deep-learning based algorithms with nerual network inference library optimized for mobile.
Image IO functions in highgui module, like cv::imread and cv::imwrite, are re-implemented using stb for smaller code size. GUI functions, like cv::imshow, are discarded.
cuda and opencl are disabled because there is no cuda on mobile, no opencl on ios, and opencl on android is slow. opencv on gpu is not suitable for real productions. Write metal on ios and opengles/vulkan on android if you need good gpu acceleration.
C++ RTTI and exceptions are disabled for minimal build on mobile platforms and webassembly build. Be careful when you write cv::Mat roi = image(roirect); :P

opencv modules included

module	comment
opencv_core	Mat, matrix operations, etc
opencv_imgproc	resize, cvtColor, warpAffine, etc
opencv_highgui	imread, imwrite
opencv_features2d	keypoint feature and matcher, etc (not included in opencv 2.x package)
opencv_photo	inpaint, etc
opencv_video	opticalflow, etc

opencv modules discarded

module	comment
opencv_androidcamera	use android Camera api instead
opencv_calib3d	camera calibration, rare uses on mobile
opencv_contrib	experimental functions, build part of the source externally if you need
opencv_dnn	very slow on mobile, try ncnn for nerual network inference on mobile
opencv_dynamicuda	no cuda on mobile
opencv_flann	feature matching, rare uses on mobile, build the source externally if you need
opencv_gapi	graph based image processing, little gain on mobile
opencv_gpu	no cuda/opencl on mobile
opencv_imgcodecs	link with opencv_highgui instead
opencv_java	wrap your c++ code with jni
opencv_js	write native code on mobile
opencv_lagacy	various good-old cv routines, build part of the source externally if you need
opencv_ml	train your ML algorithm on powerful pc or server
opencv_nonfree	the SIFT and SURF, use ORB which is faster and better
opencv_objdetect	HOG, cascade detector, use deep learning detector which is faster and better
opencv_ocl	no opencl on mobile
opencv_python	no python on mobile
opencv_shape	shape matching, rare uses on mobile, build the source externally if you need
opencv_stitching	image stitching, rare uses on mobile, build the source externally if you need
opencv_superres	do video super-resolution on powerful pc or server
opencv_ts	test modules, useless in production anyway
opencv_videoio	use android MediaCodec or ios AVFoundation api instead
opencv_videostab	do video stablization on powerful pc or server
opencv_viz	vtk is not available on mobile, write your own data visualization routines

11 KiB Raw Blame History