文件列表:

3rdpartylibs (0, 2018-11-27)
3rdpartylibs\opencv-3.4.3 (0, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include (0, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv (0, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv\cv.h (3153, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv\cv.hpp (2649, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv\cvaux.h (2523, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv\cvaux.hpp (2374, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv\cvwimage.h (2176, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv\cxcore.h (2424, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv\cxcore.hpp (2443, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv\cxeigen.hpp (2257, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv\cxmisc.h (129, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv\highgui.h (2226, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv\ml.h (2145, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv2 (0, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv2\core.hpp (150773, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv2\core (0, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv2\core\affine.hpp (21525, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv2\core\base.hpp (31336, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv2\core\bufferpool.hpp (845, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv2\core\check.hpp (6769, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv2\core\core.hpp (2366, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv2\core\core_c.h (130799, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv2\core\cuda.hpp (36890, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv2\core\cuda.inl.hpp (13916, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv2\core\cuda (0, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv2\core\cuda\block.hpp (8327, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv2\core\cuda\border_interpolate.hpp (24756, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv2\core\cuda\color.hpp (15518, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv2\core\cuda\common.hpp (3814, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv2\core\cuda\datamov_utils.hpp (4720, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv2\core\cuda\detail (0, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv2\core\cuda\detail\color_detail.hpp (222523, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv2\core\cuda\detail\reduce.hpp (15400, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv2\core\cuda\detail\reduce_key_val.hpp (23329, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv2\core\cuda\detail\transform_detail.hpp (17825, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv2\core\cuda\detail\type_traits_detail.hpp (8599, 2018-11-27)
3rdpartylibs\opencv-3.4.3\include\opencv2\core\cuda\detail\vec_distance_detail.hpp (5219, 2018-11-27)
... ...

# FaceRecognition SDK A7 edition, cross compiled by Ubuntu16.04LTS 32bit ### SDK授权 SDK采用软授权,执行`example`,可以获取机器码。调用SetLicenseFile函数,可以设置授权文件的名称。 ### 人脸比对一般流程 ```flow st=>start: 开始 op1=>operation: cv::imread,读取图片1和图片2(in BGR format) op2=>operation: VisionFaceDetect,获取图片1和图片2中的人脸位置 op3=>operation: VisionFaceAlignment,归一化图片1和图片2中的人脸 op4=>operation: VisionFaceFeature,获取图片1和图片2中的人脸特征 op5=>operation: VisionFaceFeature,比对两人脸特征,获取相似度 e=>end: 结束 st->op1->op2->op3->op4->op5->e ``` ### 1.  人脸检测 ```c++ namespace vision { class VISION_API VisionFaceDetect { public: /* 默认析构函数 */ virtual ~VisionFaceDetect() {} /* 初始化,输入参数文件param_path和所要使用的GPU编号device_id */ virtual bool Init(std::string param_path = "models/face_detect_cascade.json", int device_id = 0) = 0; /* 输入一张图像img,获取人脸信息(包括人脸框和关键点);在输入视频的情况下,将is_video_stream设置为true,能够极大提升检测速度 */ virtual std::vector GetFaces(const cv::Mat& img, bool is_video_stream = false) = 0; /* 输入一张图像img,返回所有的人脸框以及对应的人脸关键点(引用形式返回) */ virtual std::vector GetFaces(const cv::Mat& img, std::vector>& key_pts, bool is_video_stream = false) = 0; /* 获取图像img中最大人脸的人脸框 */ virtual std::vector GetMaximumFace(const cv::Mat& img, bool is_video_stream = false) = 0; /* 输入一张图像img,图像img中最大人脸的人脸框以及对应的人脸关键点(引用形式返回) */ virtual std::vector GetMaximumFace(const cv::Mat& img, std::vector>& key_pts, bool is_video_stream = false) = 0; }; /* 构造对象,默认使用GPU 0 */ VISION_API VisionFaceDetect* instantiateVisionFaceDetect(int device_id = 0); /* 销毁对象 */ VISION_API void destroyVisionFaceDetect(VisionFaceDetect* ptr); } ``` 示例代码 ```c++ // 摄像头测试 void CameraTest() { VisionFaceDetect *face_detect = instantiateVisionFaceDetect(); cv::VideoCapture capture(0); Mat img; while (true) { capture >> img; if (!img.empty()) { Mat img_show = img.clone(); double start = cv::getTickCount(); std::vector face_rects; std::vector> key_pts; face_rects = face_detect->GetFaces(img, key_pts); // 获取人脸框位置同时获取每张脸的5个关键点 for (int i = 0; i < face_rects.size(); i++) { face_detect->DrawFaceRect(img_show, face_rects.at(i)); } cv::imshow("img_show", img_show); if (cv::waitKey(5) == 27) { break; } } } vision::destroyVisionFaceDetect(face_detect); } ``` ---- ### 2.  关键点定位和归一化 ```c++ namespace vision { class VISION_API VisionFaceAlignment { public: /* 默认析构函数 */ virtual ~VisionFaceAlignment() {} /* 初始化,输入参数文件param_path和所要使用的GPU编号device_id */ virtual bool Init(std::string param_path = "models/face_align.json", int device_id = 0) = 0; /* 输入一张图像img和人脸框face_rect,输出人脸的关键点位置 */ virtual std::vector GetKeyPoints(const cv::Mat& img, const cv::Rect& face_rect) = 0; /* 批量获取人脸的关键点位置 */ virtual std::vector> GetKeyPoints(const cv::Mat& img, const std::vector& face_rects) = 0; /* 输入一张图像img和人脸框face_rect,输出一张归一化人脸(180x220)。若padding是true,则在图像变换的时候采用周边的像素进行补充 */ virtual cv::Mat GetAlignedFace(const cv::Mat& img, const cv::Rect& face_rect, bool padding = false) = 0; /* 批量获取归一化人脸 */ virtual std::vector GetAlignedFace(const cv::Mat& img, const std::vector& face_rect, bool padding = false) = 0; /* 输入一张图像img和人脸关键点key_pts,输出一张归一化人脸 */ virtual cv::Mat GetAlignedFace(const cv::Mat& img, const std::vector& key_pts, bool padding = false) = 0; /* 批量获取归一化人脸 */ virtual std::vector GetAlignedFace(const cv::Mat& img, const std::vector>& key_pts, bool padding = false) = 0; }; /* 构造对象,新增变量device_id,默认使用GPU 0 */ VISION_API VisionFaceAlignment* instantiateVisionFaceAlignment(int device_id = 0); /* 销毁对象 */ VISION_API void destroyVisionFaceAlignment(VisionFaceAlignment* ptr); } ``` 示例代码 ```c++ void CameraTest() { VisionFaceDetect* face_detect = instantiateVisionFaceDetect(); VisionFaceAlignment* face_align = instantiateVisionFaceAlignment(); cv::VideoCapture capture(0); Mat img; while (true) { capture >> img; std::vector> key_pts; std::vector face_rects = face_detect->GetFaces(img, key_pts); // warning: CPU模式下,人脸检测不会返回关键点,需要通过VisionFaceAlignment获取关键点 if (key_pts.size() == 0) { key_pts = face_align->GetKeyPoints(img, face_rects); } for (int i = 0; i < key_pts.size(); i++) { for (int j = 0; j < key_pts.at(i).size(); j++) { cv::circle(img, key_pts.at(i).at(j), 2, cv::Scalar(255, 0, 0), 2); } } std::vector norm_faces = face_align->GetAlignedFace(img, key_pts); if (norm_faces.size() > 0) { cv::imshow("norm_face", norm_faces[0]); } cv::imshow("alignment result", img); if (cv::waitKey(5) == 27) { break; } } vision::destroyVisionFaceDetect(face_detect); vision::destroyVisionFaceAlignment(face_align); } ``` ---- ### 3.  人脸特征提取与比对 ```c++ namespace vision { class VISION_API VisionFaceFeature { public: /* 默认析构函数 */ virtual ~VisionFaceFeature() {} /* 初始化,输入参数文件param_path和所要使用的GPU编号device_id */ virtual bool Init(std::string param_path = "models/face_feature.json", int device_id = 0) = 0; /* 初始化,输入参数文件param_path和所要使用的GPU编号device_id,针对V3的模型【推荐使用】 */ virtual bool InitV3(std::string param_path = "face_feature_v3.param", int device_id = 0) = 0; /* 获取输入归一化人脸图像img的特征 */ virtual std::vector GetFeature(const cv::Mat& img) = 0; /* 批量获取归一化人脸图像特征 */ virtual std::vector> GetFeature(const std::vector& imgs) = 0; /* 获取两张人脸特征的相似度 */ virtual float GetScore(const std::vector& fea1, const std::vector& fea2, bool origin_score = false) = 0; }; /* 构造对象, 默认使用GPU 0 */ VISION_API VisionFaceFeature* instantiateVisionFaceFeature(int device_id = 0); /* 销毁对象 */ VISION_API void destroyVisionFaceFeature(VisionFaceFeature* ptr); } ``` 人脸比对示例代码 ```c++ int main(int argc, char** argv) { //vision::SetLicenseFile("d:/dip/license.lic"); if (argc == 4) { std::string model_name = std::string(argv[1]); std::string image1_name = std::string(argv[2]); std::string image2_name = std::string(argv[3]); vision::VisionFaceDetect * face_detect = vision::instantiateVisionFaceDetect(); // 包含默认初始化参数,无需再次初始化 face_detect->SetMaxWidthGlobal(480); face_detect->SetMinFaceSize(80); vision::VisionFaceAlignment *face_alignment = vision::instantiateVisionFaceAlignment(); // 包含默认初始化参数,无需再次初始化 vision::VisionFaceFeature * face_feature = vision::instantiateVisionFaceFeature(); // 针对不同的使用场景,需要用特定的模型进行初始化 face_feature->InitV3(model_name); cv::Mat img1 = cv::imread(image1_name); cv::Mat img2 = cv::imread(image2_name); // Perform Face Detect First std::vector faces1 = face_detect->GetMaximumFace(img1); if (faces1.size() < 1) { std::cout << "Failed to detect a face in " << image1_name << ", exit." << std::endl; exit(-1); } else { std::cout << "Face rectangle in " << image1_name << " is: " << faces1[0].bbox << std::endl; } std::vector faces2 = face_detect->GetMaximumFace(img2); if (faces2.size() < 1) { std::cout << "Failed to detect a face in " << image2_name << ", exit." << std::endl; exit(-1); } else { std::cout << "Face rectangle in " << image2_name << " is: " << faces2[0].bbox << std::endl; } // Face Alignment cv::Mat norm_face1 = face_alignment->GetAlignedFace(img1, faces1.at(0).key_pts); cv::Mat norm_face2 = face_alignment->GetAlignedFace(img2, faces2.at(0).key_pts); // Extract Face Feature std::vector fea1 = face_feature->GetFeature(norm_face1); std::vector fea2 = face_feature->GetFeature(norm_face2); // Feature Comapre std::cout << "Similarity: " << face_feature->GetScore(fea1, fea2) << std::endl; vision::destroyVisionFaceDetect(face_detect); vision::destroyVisionFaceAlignment(face_alignment); vision::destroyVisionFaceFeature(face_feature); } else { std::string cmd = std::string(argv[0]); std::cout << "Usage: " << cmd << " model_name image1 image2" << std::endl; } return 0; } ``` ### 使用建议及注意事项 - 根据业务场景使用`SetMinFaceSize`接口设置合理的最小人脸大小,以获取更快的人脸检测速度 - 根据业务场景使用`SetMaxWidthGlobal`接口设置合理的检测图像最长边大小,以获取更快的人脸检测速度 - 在嵌入式平台使用该模块,建议同比例减小`MinFaceSize`和`MaxWidthGlobal`,减少IO,加速人脸检测 - 两张人脸的相似度结果(比分0.5以上,说明像;比分0.65以上,比较像;比分0.75以上,非常像;该值的设定具体要参照业务场景) ---- ## Contact chendd14

Copyright(c) chendd14

近期下载者：

相关文件：

评论：[我要评论] [举报此文件]

收藏者：