点击上方↑↑↑“OpenCV学堂”关注我
来源：公众号 小MoMing工作室 授权
SeetaFace6提供了人脸的11个模型，本体验用到了其中7个。
已用到：人脸检测，关键点检测，人脸识别，性别，年龄，眼睛，活体检测；
未用到：带口罩识别，人脸追踪，人脸姿态，质量评估。
SeetaFace6：https://github.com/SeetaFace6Open/indexSeetaFace文档完善，开发方便。支持戴口罩与不带口罩的人脸识别：
支持端侧与云部署，当前开源版本的三个不同模型输的特征向量维度与推理速度：
代码演示部分，废话太多了，直接上代码：
//用到seetaface的7个模块//debug 库,11个//这四个没用到//release 库,11个//这四个没用到using namespace seeta;using namespace std;using namespace cv;//提取特征bool extract_feature(Mat img,const FaceDetector& FD,const FaceLandmarker& FL,const FaceRecognizer& FR,float* feature){SeetaImageData simg;simg.height = img.rows;simg.width = img.cols;simg.channels = img.channels();simg.data = img.data;auto faces = FD.detect(simg);if (faces.size <= 0){cout << "no face detected" << endl;return false;}SeetaPointF points[5];FL.mark(simg, faces.data[0].pos, points);FR.Extract(simg, points, feature);return true;}string get_eye_status(seeta::EyeStateDetector::EYE_STATE state){if (state == seeta::EyeStateDetector::EYE_CLOSE)return "闭合";else if (state == seeta::EyeStateDetector::EYE_OPEN)return "张开";else if (state == seeta::EyeStateDetector::EYE_RANDOM)return "无法判断";elsereturn "无法判断";}string get_fas_status(seeta::FaceAntiSpoofing::Status status) {switch (status) {case seeta::FaceAntiSpoofing::REAL:return "真实人脸";case seeta::FaceAntiSpoofing::SPOOF:return "照片人脸";case seeta::FaceAntiSpoofing::FUZZY:return "无法判断";case seeta::FaceAntiSpoofing::DETECTING:return "正在检测";}return "无法判断";}int main(){string ModelPath = "E:/workspace/03-research/03-cv/face/SeeTaFaceDemo/sf3.0_models/";//1.人脸检测模型初始化ModelSetting FD_setting;FD_setting.append(ModelPath + "face_detector.csta");FD_setting.set_device(ModelSetting::CPU);FD_setting.set_id(0);FaceDetector FD(FD_setting);//2.人脸关键点模型初始化ModelSetting PD_setting;PD_setting.append(ModelPath + "face_landmarker_pts5.csta");FaceLandmarker FL(PD_setting);//3.人脸识别模型初始化ModelSetting fr_setting;fr_setting.append(ModelPath + "face_recognizer.csta");FaceRecognizer FR(fr_setting);//4.性别检测模型初始化ModelSetting gb_setting(ModelPath + "gender_predictor.csta");GenderPredictor GP(gb_setting);//5.年龄检测模型初始化ModelSetting ap_setting(ModelPath + "age_predictor.csta");AgePredictor AP(ap_setting);//6.眼睛状态模型初始化ModelSetting setting;setting.append(ModelPath + "eye_state.csta");EyeStateDetector EBD(setting);//7.活体检测模型初始化ModelSetting anti_setting;anti_setting.append(ModelPath + "fas_first.csta");anti_setting.append(ModelPath + "fas_second.csta");FaceAntiSpoofing FAS(anti_setting);FAS.SetThreshold(0.3, 0.90);//设置默认阈值，另外一组阈值为(0.7, 0.55)FAS.SetBoxThresh(0.9);//建立人脸数据库的人脸特征向量：这里只有两张人脸1.jpg(刘德华)，2.jpg(薇娅)vector<pair<string, shared_ptr<float> > > feature_db;shared_ptr<float> feature1(new float[FR.GetExtractFeatureSize()]);Mat ldh = imread("1.jpg");extract_feature(ldh,FD, FL, FR,feature1.get());feature_db.emplace_back(pair<string, shared_ptr<float>>("刘德华", feature1));shared_ptr<float> feature2(new float[FR.GetExtractFeatureSize()]);Mat wy = imread("2.jpg");extract_feature(wy,FD, FL, FR, feature2.get());feature_db.emplace_back(pair<string, shared_ptr<float>>("薇娅", feature2));namedWindow("SeetaFaceAntiSpoofing", 0);Mat frame;VideoCapture capture("F:/20201204-WY-LDH-cut.mp4");// ");VideoWriter writer;cv::resize(ldh, ldh, cv::Size(120, 160));cv::resize(wy, wy, cv::Size(120, 160));if (!capture.isOpened()){cout << "fail to open!" << endl;return -1;}while (true){if (!capture.read(frame)){cout << "can not read any frame" << endl;break;}//ImageData image = frame;SeetaImageData image;image.height = frame.rows;image.width = frame.cols;image.channels = frame.channels();image.data = frame.data;auto faces = FD.detect(image);cout << "faces.size:" << faces.size << endl;for (int i = 0; i < faces.size; i++){vector<string> labels;Scalar color(0x00, 0xA0, 0x00);//----人脸----auto face = faces.data[i].pos;//----关键点检测----vector<SeetaPointF> points(FL.number());FL.mark(image, face, points.data());//----人脸识别----unique_ptr<float[]> feature(new float[FR.GetExtractFeatureSize()]);FR.Extract(image, points.data(), feature.get());//人脸识别float threshold = 0.60;int64_t target_index = -1;float max_sim = 0;for (size_t index = 0; index < feature_db.size(); ++index){auto& pair_name_feat = feature_db[index];float current_sim = FR.CalculateSimilarity(feature.get(), pair_name_feat.second.get());if (current_sim > max_sim){max_sim = current_sim;target_index = index;}}if (max_sim > threshold)labels.push_back(feature_db[target_index].first+"（相似度:"+ to_string(max_sim *100).substr(0,5) +"）");elselabels.push_back("查无此人");//----性别----GenderPredictor::GENDER gender;GP.PredictGenderWithCrop(image, points.data(), gender);string gender_str = (string("性别：") + (gender == GenderPredictor::GENDER::MALE ? "男" : "女"));//----年龄----int age;AP.PredictAgeWithCrop(image, points.data(), age);labels.push_back(gender_str + string("，年龄：") + to_string(age));//----眼睛状态----EyeStateDetector::EYE_STATE leftstate, rightstate;EBD.Detect(image, points.data(), leftstate, rightstate);labels.push_back(string("左眼：") + get_eye_status(leftstate)+ string("，右眼：") + get_eye_status(rightstate));//活体检测auto status = FAS.Predict(image, face, points.data());//PredictVideofloat clarity;float reality;FAS.GetPreFrameScore(&clarity, &reality);labels.push_back(string("活体检测：") + get_fas_status(status));if (status == FaceAntiSpoofing::SPOOF)color = Scalar(0x00, 0x00, 0xB0);drawResult(color, labels, 0, 0.0f, face.x, face.y, face.x + face.width, face.y + face.height, frame);}Scalar title_color(0x00, 0x8C, 0xFF);//绘制人脸库int space = 6;frame(cv::Rect(frame.cols / 2 - ldh.cols - space / 2 - space, frame.rows - ldh.rows - space -space, ldh.cols * 2+space+ 2*space, ldh.rows + 2 * space)) = title_color;ldh.copyTo(frame(cv::Rect(frame.cols / 2 - ldh.cols - space/2, frame.rows -ldh.rows - space, ldh.cols, ldh.rows)));wy.copyTo(frame(cv::Rect(frame.cols / 2 + space / 2, frame.rows - wy.rows - space, wy.cols, wy.rows)));static cv::Ptr<cv::freetype::FreeType2> ft2;if (!ft2) {ft2 = cv::freetype::createFreeType2();ft2->loadFontData("c:/windows/fonts/msyh.ttc", 0);}int baseline = 0;int fontHeight = 45;cv::Size text_size;if (ft2) text_size = ft2->getTextSize("人脸库(2张)", fontHeight, -1, &baseline);if (ft2)ft2->putText(frame, "人脸库(2张)", cv::Point(frame.cols/2 - text_size.width/2, frame.rows - ldh.rows/2), fontHeight, title_color, -1, 16, true);fontHeight = 60;//绘制最上面的titlestring title = "SeetaFace人脸库体验";text_size = ft2->getTextSize(title, fontHeight, -1, &baseline);ft2->putText(frame, title, cv::Point(frame.cols / 2 - text_size.width / 2, 60), fontHeight, title_color, -1, 16, true);fontHeight = 30;title = "用到：人脸,关键点,识别,性别,年龄,眼睛,活体；未用到：口罩,追踪,姿态,质量评估";text_size = ft2->getTextSize(title, fontHeight, -1, &baseline);ft2->putText(frame, title, cv::Point(frame.cols / 2 - text_size.width / 2, 60+50), fontHeight, title_color, -1, 16, true);//写入文件if (!writer.isOpened()) {writer.open("F:/setaface.avi", VideoWriter::fourcc('M', 'J', 'P', 'G'), 30, cv::Size(frame.cols, frame.rows), true);}if (writer.isOpened()) {writer.write(frame);}//窗口显示imshow("SeetaFaceAntiSpoofing", frame);//Esc键退出if (waitKey(1) == 27)break;}writer.release();capture.release();return 0;}using namespace cv;using namespace std;//float colors[6][3] = { {1,0,1}, {0,0,1},{0,1,1},{0,1,0},{1,1,0},{1,0,0} };////float get_color(int c, int x, int max)//{// float ratio = ((float)x / max) * 5;// int i = floor(ratio);// int j = ceil(ratio);// ratio -= i;// float r = (1 - ratio) * colors[i][c] + ratio * colors[j][c];// return r;//}void drawResult(Scalar color,std::vector<string> labels,int classId, float conf, int left, int top, int right, int bottom, Mat& frame){/*static std::vector<string> labels = {"船名: GOLDEN MONTERREY","呼号：V7MG6","MMSI: 538006355","航首向: 110度","航迹向: 108度","航速度: 11.9","经纬度：120.35889, 36.017105","更新时间: 2020-11-06 10:38:48"};*///std::string label = /*classes[classId]*/ +"(" + std::to_string(int(conf * 100)) + "%)";//classId = 0;//选颜色//static int classes = 80;//static int offset = (classId + 2) * 123457 % classes;//static float red = get_color(2, offset, classes);//static float green = get_color(1, offset, classes);//static float blue = get_color(0, offset, classes);//static cv::Scalar color(int(red * 256), int(green * 256), int(blue * 256));static cv::Ptr<cv::freetype::FreeType2> ft2;if (!ft2) {ft2 = cv::freetype::createFreeType2();ft2->loadFontData("c:/windows/fonts/msyh.ttc", 0);}int fontHeight = 25;int fontTotalHeight = fontHeight * labels.size();int thickness = -1;int linestyle = LineTypes::LINE_AA;int baseline = 0;int max_label_index = 0;int padding = 5;for (int i = 1; i < labels.size(); i++) {if (labels[i].length() > labels[max_label_index].length())max_label_index = i;}cv::Size text_size;if (ft2) text_size = ft2->getTextSize(labels[max_label_index], fontHeight, thickness, &baseline);else text_size = cv::getTextSize(labels[max_label_index], fontHeight, 1.0f, thickness = 0, &baseline);fontTotalHeight += 2 * padding + 2* labels.size();text_size.width += 2 * padding;cv::Point pt1, pt2;cv::Point pt_text_bg1, pt_text_bg2;cv::Point pt_text;//物体框pt1.x = left;pt1.y = top;pt2.x = right;pt2.y = bottom;//文本背景框pt_text_bg1.x = left;pt_text_bg1.y = top - fontTotalHeight;pt_text_bg2.x = std::max(left + text_size.width, right);pt_text_bg2.y = top;//文本原点(左下角)pt_text.x = left + padding;pt_text.y = top - padding;static int rect_line_width = 2;//std::max(1.0f, show_img->rows * .002f);cv::rectangle(frame, pt1, pt2, color, rect_line_width, linestyle, 0);cv::rectangle(frame, pt_text_bg1, pt_text_bg2, color, rect_line_width, linestyle, 0);cv::rectangle(frame, pt_text_bg1, pt_text_bg2, color, cv::FILLED, linestyle, 0);static cv::Scalar text_color = CV_RGB(255, 255, 255);for (int i = labels.size() - 1; i >= 0; i--) {if (ft2)ft2->putText(frame, labels[i], pt_text, fontHeight+(i==0?5:0), text_color, thickness, linestyle, true);else putText(frame, labels[i], pt_text, fontHeight, 1.0f, text_color, thickness = 0, linestyle, true);pt_text.y -= (fontHeight + 2);//((labels.size()-1 -i) * fontHeight);}}
视频B站观看地址 ：
https://www.bilibili.com/video/BV1qK411u7TF/运行效果视频截图：
推荐阅读
Tensorflow + OpenCV4 安全帽检测模型训练与推理
OpenCV Python + Tesseract-OCR轻松实现中文识别
YOLOv5在最新OpenVINO 2021R02版本的部署与代码演示详解