如何在OpenHarmony上使用SeetaFace2人脸识别库？

简介

相信大部分同学们都已了解或接触过OpenAtom OpenHarmony（以下简称“OpenHarmony”）了，但你一定没在OpenHarmony上实现过人脸识别功能，跟着本文带你快速在OpenHarmony标准设备上基于SeetaFace2和OpenCV实现人脸识别。

项目效果

本项目实现了导入人脸模型、人脸框选和人脸识别三大功能，操作流程如下：

1. 录入页面点击右下角按钮，跳转拍摄页面进行拍照；

2. 选择一张或多张人脸作为训练模型，并设置对应的名字；

3. 选择一张未录入的人脸图片，点击框选按钮实现人脸图片框选功能；

4. 最后点击识别，应用会对当前图片进行匹配，最终在界面中显示识别结果。

快速上手

设备端开发

设备端通过OpenCV对图像进行处理并通过Seetaface2对图形数据进行人脸头像的识别，最终输出对应的NAPI接口提供给应用端调用。因此设备端开发主要涉及到OpenCV和Seetaface2的移植以及NAPI接口的开发。

OpenCV库移植

OpenCV是一个功能非常强大的开源计算机视觉库。此库已由知识体系工作组移植到了OpenHarmony中，后期还会将此库合入到主仓。在此库上主仓之前，我们只需要以下几个步骤就可以实现OpenCV的移植使用。

1. 通过以下命令下载已经移植好的OpenCV

git clone git@gitee.com:zhong-luping/ohos_opencv.git

2. 将OpenCV拷贝到OpenHarmony目录的third_party下

cp -raf opencv ~/openharmony/third_party/

3. 适当裁剪编译选项

打开OpenCV目录下的BUILD.gn，如下：

不需要video以及flann功能，将对应的模块注释即可。

import("//build/ohos.gni")
group("opencv") {deps = ["//third_party/opencv/modules/core:opencv_core",//  "//third_party/opencv/modules/flann:opencv_flann","//third_party/opencv/modules/imgproc:opencv_imgproc","//third_party/opencv/modules/ml:opencv_ml","//third_party/opencv/modules/photo:opencv_photo","//third_party/opencv/modules/dnn:opencv_dnn","//third_party/opencv/modules/features2d:opencv_features2d","//third_party/opencv/modules/imgcodecs:opencv_imgcodecs","//third_party/opencv/modules/videoio:opencv_videoio","//third_party/opencv/modules/calib3d:opencv_calib3d","//third_party/opencv/modules/highgui:opencv_highgui","//third_party/opencv/modules/objdetect:opencv_objdetect","//third_party/opencv/modules/stitching:opencv_stitching","//third_party/opencv/modules/ts:opencv_ts",//   "//third_party/opencv/modules/video:opencv_video","//third_party/opencv/modules/gapi:opencv_gapi",]
}

4. 添加依赖子系统的part_name，编译框架子系统会将编译出的库拷贝到系统文件中。

此项目中我们新建了一个SeetaFaceApp的子系统，该子系统中命名part_name为SeetafaceApi，所以我们需要在对应模块中的BUILD.gn中加上part_name=“SeetafaceApi”

以module/core为例：

ohos_shared_library("opencv_core"){sources = [ ... ]
configs = [  ... ]
deps = [ ... ]
part_name = "SeetafaceApi"
}

5. 编译工程需要添加OpenCV的依赖。

在生成NAPI的BUILD.gn中添加以下依赖：

deps += [ "//third_party/opencv:opencv" ]

至此，人脸识别中OpenCV的移植使用完成。

SeetaFace2库移植

SeetaFace2是中科视拓开源的第二代人脸识别库。包括了搭建一套全自动人脸识别系统所需的三个核心模块，即：人脸检测模块FaceDetector、面部关键点定位模块FaceLandmarker以及人脸特征提取与比对模块 FaceRecognizer。

关于SeetaFace2的移植请参照文档:SeetaFace2移植开发文档。

NAPI接口开发

关于OpenHarmony中的NAPI开发，参考视频：

OpenHarmony中napi的开发视频教程。本文将重点讲解NAPI接口如何实现OpenCV以及SeetaFace的调用。

1. 人脸框获取的NAPI接口的实现。

int GetRecognizePoints(const char *image_path)；

此接口主要是通过应用层输入一张图片，通过OpenCV的imread接口获取到图片数据，并通过人脸检测模块FaceDetector分析获得图片中所有的人脸矩形框（矩形框是以x,y,w,h的方式）并将人脸框矩形以数组的方式返回到应用层。

人脸框矩形获取的主要代码如下：

static int RecognizePoint(string image_path, FaceRect *rect, int num)
{if (rect == nullptr) {cerr << "NULL POINT!" << endl;LOGE("NULL POINT! \n");return -1;}seeta::ModelSetting::Device device = seeta::ModelSetting::CPU;int id = 0;/* 设置人脸识别模型。*/seeta::ModelSetting FD_model( "/system/usr/model/fd_2_00.dat", device, id );seeta::ModelSetting FL_model( "/system/usr/model/pd_2_00_pts81.dat", device, id );seeta::FaceDetector FD(FD_model);seeta::FaceLandmarker FL(FL_model);FD.set(seeta::FaceDetector::PROPERTY_VIDEO_STABLE, 1);/* 读取图片数据 */auto frame = imread(image_path);seeta::cv::ImageData simage = frame;if (simage.empty()) {cerr << "Can not open image: " << image_path << endl;LOGE("Can not open image: %{public}s", image_path.c_str());return -1;}/* 图片数据进行人脸识别处理 ，获取所有的人脸框数据对象*/auto faces = FD.detect(simage);if (faces.size <= 0) {cerr << "detect " << image_path << "failed!" << endl;LOGE("detect image: %s failed!", image_path.c_str());return -1;}for (int i = 0; (i < faces.size && i < num); i++) {/* 将所有人脸框对象数据以坐标形式输出*/auto &face = faces.data[i];memcpy(&rect[i], &(face.pos), sizeof(FaceRect));}return faces.size;
}

其中FD_model是人脸检测模型，而FL_model是面部关键点定位模型(此模型分为5点定位和81点定位，本项目中使用的是81点定位模型)，这些模型从开源项目中免费获取。

通过以上方式获取到对应的人脸矩形框后，再将矩形框以数组的方式返回到应用端：

string image = path;
p = (FaceRect *)malloc(sizeof(FaceRect) * MAX_FACE_RECT);
/* 根据图片进行人脸识别并获取人脸框坐标点 */
int retval = RecognizePoint(image, p, MAX_FACE_RECT);
if (retval <= napi_ok) {LOGE("GetNapiValueString failed!");free(p);return result;
} 
/*将所有坐标点以数组方式返回到应用端*/
for (int i = 0; i < retval; i++) {int arry_int[4] = {p[i].x, p[i].y, p[i].w, p[i].h};int arraySize = (sizeof(arry_int) / sizeof(arry_int[0]));for (int j = 0; j < arraySize; j++) {napi_value num_val;if (napi_create_int32(env, arry_int[j], &num_val) != napi_ok) {LOGE("napi_create_int32 failed!");return result;}napi_set_element(env, array, i*arraySize + j, num_val);}
}
if (napi_create_object(env, &result) != napi_ok) {LOGE("napi_create_object failed!");free(p);return result;
}
if (napi_set_named_property(env, result, "recognizeFrame", array) != napi_ok) {LOGE("napi_set_named_property failed!");free(p);return result;
}
LOGI("");
free(p);
return result;

其中array是通过napi_create_array创建的一个NAPI数组对象，通过 napi_set_element将所有的矩形框数据保存到array对象中，最后通过 napi_set_named_property将array转换成应用端可识别的对象类型result并将其返回。

2. 人脸搜索识别初始化与逆初始化。

1. int FaceSearchInit()；

2. int FaceSearchDeinit()；

这2个接口主要是提供给人脸搜索以及识别调用的，初始化主要包含模型的注册以及识别模块的初始化：

static  int FaceSearchInit(FaceSearchInfo *info)
{if (info == NULL) {info = (FaceSearchInfo *)malloc(sizeof(FaceSearchInfo));if (info == nullptr) {cerr << "NULL POINT!" << endl;return -1;}}seeta::ModelSetting::Device device = seeta::ModelSetting::CPU;int id = 0;seeta::ModelSetting FD_model( "/system/usr/model/fd_2_00.dat", device, id );seeta::ModelSetting PD_model( "/system/usr//model/pd_2_00_pts5.dat", device, id );seeta::ModelSetting FR_model( "/system/usr/model/fr_2_10.dat", device, id );info->engine = make_shared<seeta::FaceEngine>(FD_model, PD_model, FR_model, 2, 16);info->engine->FD.set( seeta::FaceDetector::PROPERTY_MIN_FACE_SIZE, 80);info->GalleryIndexMap.clear();return 0;
}

而逆初始化就是做一些内存的释放。

static void FaceSearchDeinit(FaceSearchInfo *info, int need_delete)
{if (info != nullptr) {if (info->engine != nullptr) {}info->GalleryIndexMap.clear();if (need_delete) {free(info);info = nullptr;}}
}

3. 人脸搜索识别注册接口的实现。

int FaceSearchRegister(const char *value);

需要注意的是，该接口需要应用端传入一个json数据的参数，主要包含注册人脸的名字，图片以及图片个数，如{“name”:“刘德华”,“sum”:“2”,“image”:{“11.jpg”,“12.jpg”}}。而解析参数的时候需要调用 napi_get_named_property对json数据的各个对象进行解析，具体代码如下：

napi_get_cb_info(env, info, &argc, &argv, &thisVar, &data);
napi_value object = argv;
napi_value value = nullptr;if (napi_get_named_property(env, object, (const char *)"name", &value) == napi_ok) {char name[64] = {0};if (GetNapiValueString(env, value, (char *)name, sizeof(name)) < 0) {LOGE("GetNapiValueString failed!");return result;}reg_info.name = name;
}
LOGI("name = %{public}s", reg_info.name.c_str());
if (napi_get_named_property(env, object, (const char *)"sum", &value) == napi_ok) {if (napi_get_value_uint32(env, value, &sum) != napi_ok) {LOGE("napi_get_value_uint32 failed!");return result;}
}
LOGI("sum = %{public}d", sum);
if (napi_get_named_property(env, object, (const char *)"image", &value) == napi_ok) {bool res = false;if (napi_is_array(env, value, &res) != napi_ok || res == false) {LOGE("napi_is_array failed!");return result;}for (int i = 0; i < sum; i++) {char image[256] = {0};napi_value imgPath = nullptr;if (napi_get_element(env, value, i, &imgPath) != napi_ok) {LOGE("napi_get_element failed!");return result;}if (GetNapiValueString(env, imgPath, (char *)image, sizeof(image)) < 0) {LOGE("GetNapiValueString failed!");return result;}reg_info.path = image;if (FaceSearchRegister(g_FaceSearch, reg_info) != napi_ok) {retval = -1;break;}}
}

通过napi_get_cb_info获取从应用端传来的参数，并通过 napi_get_named_property获取对应的name以及图片个数，最后通过napi_get_element获取图片数组中的各个image，将name和image通过FaceSearchRegister接口将图片和名字注册到SeetaFace2模块的识别引擎中。具体实现如下：

static int FaceSearchRegister(FaceSearchInfo &info, RegisterInfo &gegister)
{if (info.engine == nullptr) {cerr << "NULL POINT!" << endl;return -1;}seeta::cv::ImageData image = cv::imread(gegister.path);auto id = info.engine->Register(image);if (id >= 0) {info.GalleryIndexMap.insert(make_pair(id, gegister.name));}return 0;
}

注册完数据后，后续可以通过该引擎来识别对应的图片。

4. 获取人脸搜索识别结果接口的实现。

char *FaceSearchGetRecognize(const char *image_path);

该接口实现了通过传入一张图片，在识别引擎中进行搜索识别。如果识别引擎中有类似的人脸注册，则返回对应人脸注册时的名字，否则返回不识别(ignored)字样。该方法是通过异步回调的方式实现的：

// 创建async work，创建成功后通过最后一个参数(commandStrData->asyncWork)返回async work的handle
napi_value resourceName = nullptr;
napi_create_string_utf8(env, "FaceSearchGetPersonRecognizeMethod", NAPI_AUTO_LENGTH, &resourceName);
napi_create_async_work(env, nullptr, resourceName, FaceSearchRecognizeExecuteCB, FaceSearchRecognizeCompleteCB,(void *)commandStrData, &commandStrData->asyncWork);// 将刚创建的async work加到队列，由底层去调度执行
napi_queue_async_work(env, commandStrData->asyncWork);

其中FaceSearchRecognizeExecuteCB实现了人脸识别

static void FaceSearchRecognizeExecuteCB(napi_env env, void *data)
{CommandStrData *commandStrData = dynamic_cast<CommandStrData*>((CommandStrData *)data);if (commandStrData == nullptr) {HILOG_ERROR("nullptr point!", __FUNCTION__, __LINE__);return;}FaceSearchInfo faceSearch = *(commandStrData->mFaceSearch);commandStrData->result = FaceSearchSearchRecognizer(faceSearch, commandStrData->filename);LOGI("Recognize result : %s !", __FUNCTION__, __LINE__, commandStrData->result.c_str());
}

FaceSearchRecognizeCompleteCB函数通过napi_resolve_deferred接口将识别结果返回到应用端。

static void FaceSearchRecognizeCompleteCB(napi_env env, napi_status status, void *data)
{CommandStrData *commandStrData = dynamic_cast<CommandStrData*>((CommandStrData *)data);napi_value result;if (commandStrData == nullptr || commandStrData->deferred == nullptr) {LOGE("nullptr", __FUNCTION__, __LINE__);if (commandStrData != nullptr) {napi_delete_async_work(env, commandStrData->asyncWork);delete commandStrData;}return;}const char *result_str = (const char *)commandStrData->result.c_str();if (napi_create_string_utf8(env, result_str, strlen(result_str), &result) != napi_ok) {LOGE("napi_create_string_utf8 failed!", __FUNCTION__, __LINE__);napi_delete_async_work(env, commandStrData->asyncWork);delete commandStrData;return;}napi_resolve_deferred(env, commandStrData->deferred, result);napi_delete_async_work(env, commandStrData->asyncWork);delete commandStrData;
}

通过人脸特征提取与比对模块，对传入的数据与已注册的数据进行对比，并通过返回对比的相似度来进行判断当前人脸是否为可识别的，最后返回识别结果。具体实现代码：

static string FaceSearchSearchRecognizer(FaceSearchInfo &info, string filename)
{if (info.engine == nullptr) {cerr << "NULL POINT!" << endl;return "recognize error 0";}string name;float threshold = 0.7f;seeta::QualityAssessor QA;auto frame = cv::imread(filename);if (frame.empty()) {LOGE("read image %{public}s failed!", filename.c_str());return "recognize error 1!";}seeta::cv::ImageData image = frame;std::vector<SeetaFaceInfo> faces = info.engine->DetectFaces(image);for (SeetaFaceInfo &face : faces) {int64_t index = 0;float similarity = 0;auto points = info.engine->DetectPoints(image, face);auto score = QA.evaluate(image, face.pos, points.data());if (score == 0) {name = "ignored";} else {auto queried = info.engine->QueryTop(image, points.data(), 1, &index, &similarity);// no face queried from databaseif (queried < 1) continue;// similarity greater than threshold, means recognizedif( similarity > threshold ) {name = info.GalleryIndexMap[index];}}}LOGI("name : %{public}s \n", name.length() > 0 ? name.c_str() : "null");return name.length() > 0 ? name : "recognize failed";
}

至此，所有的NAPI接口已经开发完成。

5. NAPI库编译开发完NAPI接口后，我们需要将我们编写的库加入到系统中进行编译，我们需要添加一个自己的子系统。

首先在库目录下新建一个ohos.build

{"subsystem": "SeetafaceApp","parts": {"SeetafaceApi": {"module_list": ["//seetaface:seetafaceapp_napi"],"test_list": [ ]}}
}

其次同一目录新建一个BUILD.gn，将库源文件以及对应的依赖加上，如下：

import("//build/ohos.gni")config("lib_config") {cflags_cc = ["-frtti","-fexceptions","-DCVAPI_EXPORTS","-DOPENCV_ALLOCATOR_STATS_COUNTER_TYPE=int","-D_USE_MATH_DEFINES","-D__OPENCV_BUILD=1","-D__STDC_CONSTANT_MACROS","-D__STDC_FORMAT_MACROS","-D__STDC_LIMIT_MACROS","-O2","-Wno-error=header-hygiene",]
}ohos_shared_library("seetafaceapp_napi") {sources = ["app.cpp",]include_dirs = ["./","//third_party/opencv/include","//third_party/opencv/common","//third_party/opencv/modules/core/include","//third_party/opencv/modules/highgui/include","//third_party/opencv/modules/imgcodecs/include","//third_party/opencv/modules/imgproc/include","//third_party/opencv/modules/calib3d/include","//third_party/opencv/modules/dnn/include","//third_party/opencv/modules/features2d/include","//third_party/opencv/modules/flann/include","//third_party/opencv/modules/ts/include","//third_party/opencv/modules/video/include","//third_party/opencv/modules/videoio/include","//third_party/opencv/modules/ml/include","//third_party/opencv/modules/objdetect/include","//third_party/opencv/modules/photo/include","//third_party/opencv/modules/stitching/include","//third_party/SeetaFace2/FaceDetector/include","//third_party/SeetaFace2/FaceLandmarker/include","//third_party/SeetaFace2/FaceRecognizer/include","//third_party/SeetaFace2/QualityAssessor/include","//base/accessibility/common/log/include","//base/hiviewdfx/hilog_lite/interfaces/native/innerkits"]deps = [ "//foundation/ace/napi:ace_napi" ]deps += [ "//third_party/opencv:opencv" ]deps += [ "//third_party/SeetaFace2:SeetaFace2" ]external_deps = ["hiviewdfx_hilog_native:libhilog",]configs = [":lib_config"]# 指定库生成的路径relative_install_dir = "module"# 子系统及其组件，后面会引用subsystem_name = "SeetafaceApp"part_name = "SeetafaceApi"
}

添加完对应的文件后，我们需要将我们的子系统添加到系统中进行编译，打开build/subsystem_config.json并在最后添加以下代码：

"SeetafaceApp": {"path": "seetaface","name": "SeetafaceApp"
}

添加完子系统再修改产对应的品配置

打开productdefine/common/products/rk3568.json并在最后添加以下代码：

"SeetafaceApp:SeetafaceApi":{}

做完以上修改后我们就可以通过以下命令直接编译NAPI的库文件了：

./build.sh --product-name rk3568 --ccache

参考RK3568快速上手-镜像烧录完成烧录即可。

应用端开发

在完成设备NAPI功能开发后，应用端通过调用NAPI组件中暴露给应用的人脸识别接口，即可实现对应功能。接下来就带着大家使用NAPI实现人脸识别功能。

开发准备

1. 下载DevEco Studio 3.0 Beta4；

2. 搭建开发环境，参考开发准备；

3. 了解属性eTS开发，参考eTS语言快速入门；

SeetaFace2初始化

1. 首先将SeetaFace2 NAPI接口声明文件放置于SDK目录/api下；

2. 然后导入SeetaFace2 NAPI模块；ck-start/star

3. 调用初始化接口；

// 首页实例创建后
async aboutToAppear() {await StorageUtils.clearModel();CommonLog.info(TAG,'aboutToAppear')// 初始化人脸识别let res = SeetafaceApp.FaceSearchInit()CommonLog.info(TAG,`FaceSearchInit res=${res}`)this.requestPermissions()
}// 请求权限
requestPermissions(){CommonLog.info(TAG,'requestPermissions')let context = featureAbility.getContext()context.requestPermissionsFromUser(PERMISSIONS, 666,(res)=>{this.getMediaImage()})
}

获取所有人脸图片

通过文件管理模块fileio和媒体库管理mediaLibrary，获取指定应用数据目录下所有的图片信息，并将路径赋值给faceList，faceList数据用于Image组件提供url进行加载图片

// 获取所有图片
async getMediaImage(){let context = featureAbility.getContext();// 获取本地应用沙箱路径let localPath = await context.getOrCreateLocalDir()CommonLog.info(TAG, `localPath:${localPath}`)let facePath = localPath + "/files"// 获取所有照片this.faceList = await FileUtil.getImagePath(facePath)
}

设置人脸模型

获取选中的人脸图片地址和输入的名字，调用SeetafaceApp.FaceSearchRegister(params)进行设置人脸模型。其中参数params由name名字、image图片地址集合和sum图片数量组成。

async submit(name) {if (!name || name.length == 0) {CommonLog.info(TAG, 'name is empty')return}let selectArr = this.faceList.filter(item => item.isSelect)if (selectArr.length == 0) {CommonLog.info(TAG, 'faceList is empty')return}// 关闭弹窗this.dialogController.close()try {let urls = []let files = []selectArr.forEach(item => {let source = item.url.replace('file://', '')CommonLog.info(TAG, `source:${source}`)urls.push(item.url)files.push(source)})// 设置人脸识别模型参数let params = {name: name,image: files,sum: files.length}CommonLog.info(TAG, 'FaceSearchRegister' + JSON.stringify(params))let res = SeetafaceApp.FaceSearchRegister(params)CommonLog.info(TAG, 'FaceSearchRegister res ' + res)// 保存已设置的人脸模型到轻量存储let data = {name:name,urls:urls}let modelStr = await StorageUtils.getModel()let modelList = JSON.parse(modelStr)modelList.push(data)StorageUtils.setModel(modelList)router.back()} catch (err) {CommonLog.error(TAG, 'submit fail ' + err)}
}

实现框选人脸

调用SeetafaceApp.GetRecognizePoints传入当前图片地址，获取到人脸左上和右下坐标，再通过CanvasRenderingContext2D对象绘画出人脸框。

实现人脸识别

调用SeetafaceApp.FaceSearchGetRecognize(url)，传入图片地址对人脸进行识别并返回对应识别出来的名字。

// 人脸识别
recognize(){SeetafaceApp.FaceSearchGetRecognize(this.url).then(res=>{CommonLog.info(TAG,'recognize suceess' + JSON.stringify(res))if(res && res != 'ignored' && res != "recognize failed" && res != 'recognize error 1!'){// 赋值识别到的人物模型this.name = res}else{this.name = '未识别到该模型'}}).catch(err=>{CommonLog.error(TAG,'recognize' + err)this.name = '未识别到该模型'})
}

为了帮助到大家能够更有效的学习OpenHarmony 开发的内容，下面特别准备了一些相关的参考学习资料：