文章目錄

(資料圖片僅供參考)

前言一、face_recognition1.1 安裝1.2 檢測人臉位置1.3 識別人臉 二、PaddleDetection2.1 安裝2.2 運(yùn)行 三、DeepFace3.1 安裝3.2 檢測人臉位置3.3 人臉屬性分析 四、insightface4.1 安裝4.2 運(yùn)行 五、SeetaFaceEngine5.1 編譯5.2 人臉檢測5.3 face alignment5.4 人臉檢測相似率 六、OpenFace6.1 安裝6.2 運(yùn)行 參考

前言

人臉識別是機(jī)器學(xué)習(xí)熱門領(lǐng)域之一，在 github 上有很多項(xiàng)目實(shí)現(xiàn)了各種人臉識別功能，以下面6個測試軟件使用

https://github.com/ageitgey/face_recognitionhttps://github.com/PaddlePaddle/PaddleDetectionhttps://github.com/serengil/deepfacehttps://github.com/deepinsight/insightfacehttps://github.com/seetaface/SeetaFaceEnginehttps://github.com/TadasBaltrusaitis/OpenFace

一、face_recognition

face_recognition 是世界上最簡潔的人臉識別庫，可以使用 Python 和命令行工具提取、識別、操作人臉。

face_recognition 項(xiàng)目的人臉識別是基于業(yè)內(nèi)領(lǐng)先的C++開源庫 dlib中的深度學(xué)習(xí)模型，用Labeled Faces in the Wild人臉數(shù)據(jù)集進(jìn)行測試，有高達(dá)99.38%的準(zhǔn)確率。但對小孩和亞洲人臉的識別準(zhǔn)確率尚待提升。

1.1 安裝

pip install face_recognition

1.2 檢測人臉位置

可以使用命令行命令 face_detection來識別人臉，下面以胡歌照片為例，來演示具體使用

face_detection faces/huge.jpg# 輸出：faces/huge.jpg,101,221,173,149

使用命令行只顯示了位置的具體坐標(biāo)，不能準(zhǔn)確的用肉眼查看，可以使用 python 來標(biāo)記

import face_recognitionfrom PIL import Image, ImageDrawimage = face_recognition.load_image_file("huge.jpg")face_locations = face_recognition.face_locations(image)pil_image = Image.fromarray(image)draw = ImageDraw.Draw(pil_image)for (top, right, bottom, left) in face_locations:    draw.rectangle(((left, top), (right, bottom)), outline=(0, 0, 255))del drawpil_image.save("huge_face.jpg")

多張人臉檢測

原始照片

標(biāo)記人臉位置照片

1.3 識別人臉

face_recognition 不僅支持識別人臉?biāo)谡掌恢?，更能識別人臉?biāo)淼娜?/p>

將 [‘劉詩詩.jpg’, ‘唐嫣.jpg’, ‘楊冪.jpg’, ‘胡歌.jpg’, ‘霍建華.jpg’, ‘黃志瑋.jpg’] 照片放在一個文件夾下，例如我的是 known 文件夾下，再將仙劍三海報(bào) all.jpg 放在和腳本同一目錄下，開始識別人臉

測試的6張照片都是從網(wǎng)上找的，鏈接如下

劉詩詩唐嫣楊冪胡歌霍建華黃志瑋

import face_recognitionimport osfrom PIL import Image, ImageDraw, ImageFontimport numpy as npfont = ImageFont.truetype("C:\\Windows\\Fonts\\simsun.ttc", 40, encoding="utf-8")known_path = "../known"known_face_names = []known_face_encodings = []images = os.listdir(known_path)print(images)for image in images:    if image.endswith("jpg"):        known_face_names.append(os.path.basename(image).split(".")[0])        image_data = face_recognition.load_image_file(os.path.join(known_path, image))        known_face_encodings.append(face_recognition.face_encodings(image_data)[0])all_face_path = "all.jpg"all_image = face_recognition.load_image_file(all_face_path)all_face_locations = face_recognition.face_locations(all_image)all_face_encodings = face_recognition.face_encodings(all_image, all_face_locations)pil_image = Image.fromarray(all_image)draw = ImageDraw.Draw(pil_image)for (top, right, bottom, left), face_encoding in zip(all_face_locations, all_face_encodings):    matches = face_recognition.compare_faces(known_face_encodings, face_encoding, tolerance=0.5)    name = "未知"    face_distances = face_recognition.face_distance(known_face_encodings, face_encoding)    best_match_index = np.argmin(face_distances)    if matches[best_match_index]:        name = known_face_names[best_match_index]    draw.rectangle(((left, top), (right, bottom)), outline=(0, 0, 255))    text_width, text_height = draw.textsize(name, font=font)    draw.text((left + 6, bottom - text_height - 5), name, fill=(255, 255, 255, 255), font=font)del drawpil_image.save("all_faces.jpg")

二、PaddleDetection

PaddleDetection為基于飛槳 PaddlePaddle 的端到端目標(biāo)檢測套件，內(nèi)置30+模型算法及250+預(yù)訓(xùn)練模型，覆蓋目標(biāo)檢測、實(shí)例分割、跟蹤、關(guān)鍵點(diǎn)檢測等方向，其中包括服務(wù)器端和移動端高精度、輕量級產(chǎn)業(yè)級SOTA模型、冠軍方案和學(xué)術(shù)前沿算法，并提供配置化的網(wǎng)絡(luò)模塊組件、十余種數(shù)據(jù)增強(qiáng)策略和損失函數(shù)等高階優(yōu)化支持和多種部署方案，在打通數(shù)據(jù)處理、模型開發(fā)、訓(xùn)練、壓縮、部署全流程的基礎(chǔ)上，提供豐富的案例及教程，加速算法產(chǎn)業(yè)落地應(yīng)用。

2.1 安裝

下載源碼，根據(jù)readme安裝，注意下載源碼版本需要根 paddlepaddle 版本對應(yīng)。

安裝過程中，安裝cython bbox 失敗，解決方法：windows下安裝cython-bbox失敗。下載資源：cython+bbox-0.1.3

2.2 運(yùn)行

PaddleDetection 內(nèi)置一個高效、高速的人臉檢測解決方案，包括最先進(jìn)的模型和經(jīng)典模型

python tools/infer.py -c configs/face_detection/blazeface_1000e.yml -o weights=https://paddledet.bj.bcebos.com/models/blazeface_1000e.pdparams --infer_img=C:\Users\supre\Desktop\faces\all.jpg --output_dir=infer_output/ --draw_threshold=0.6

三、DeepFace

Deepface 是一個輕量級的人臉面部識別和面部屬性分析（年齡、性別、情感和種族）框架。它是一個混合的人臉識別框架，包裝了最先進(jìn)的模型：VGG-Face, Google FaceNet, OpenFace, Facebook DeepFace, DeepID, ArcFace, Dliband SFace.

實(shí)驗(yàn)表明，人類在面部識別任務(wù)上的準(zhǔn)確率為97.53%，而這些模型已經(jīng)達(dá)到并通過了這個準(zhǔn)確率水平。

3.1 安裝

pip install deepface

3.2 檢測人臉位置

from deepface import DeepFacefrom deepface.detectors import FaceDetectorimport cv2img_path = "C:\\Users\\supre\\Desktop\\faces\\all.jpg"detector_name = "opencv"img = cv2.imread(img_path)detector = FaceDetector.build_model(detector_name) #set opencv, ssd, dlib, mtcnn or retinafaceobj = FaceDetector.detect_faces(detector, detector_name, img)faces = []regions = []for o in obj:    face, region = o    faces.append(face)    regions.append(region)for (x, y, w, h) in regions:    cv2.rectangle(img, (x, y), (x+w, y + h), (0, 0, 255), 2)cv2.imwrite("all_deep_face.jpg", img)cv2.imshow("faces", img)cv2.waitKey(0)print("there are ",len(obj)," faces")

3.3 人臉屬性分析

運(yùn)行下面代碼會從 github 下載訓(xùn)練好的模型文件，如果下載太慢可手動下載：https://github.com/serengil/deepface_models/releases/

from deepface import DeepFaceobj = DeepFace.analyze(img_path = "faces/huge.jpg",         actions = ["age", "gender", "race", "emotion"])print(obj)

輸出：

{"age": 31, "region": {"x": 141, "y": 90, "w": 92, "h": 92}, "gender": "Man", "race": {"asian": 86.62416855240873, "indian": 0.2717677898641103, "black": 0.025535856615095234, "white": 11.001530200334203, "middle eastern": 0.36970814565319693, "latino hispanic": 1.707288910883004}, "dominant_race": "asian", "emotion": {"angry": 4.005255788877951, "disgust": 1.1836746688898558e-05, "fear": 91.75890038960578, "happy": 1.023393651002267, "sad": 0.9277909615809299, "surprise": 2.081933555420253, "neutral": 0.20271948350039026}, "dominant_emotion": "fear"}

四、insightface

insightface 是一個開源的二維和三維深度面部分析工具箱，主要基于 PyTorch 和 MXNet。實(shí)現(xiàn)了很多人臉識別、人臉檢測和人臉對齊算法，為訓(xùn)練和部署進(jìn)行了優(yōu)化。

4.1 安裝

pip install insightface

4.2 運(yùn)行

運(yùn)行出現(xiàn)報(bào)錯：TypeError: __init__() got an unexpected keyword argument "provider_options"

查找資料Error “got an unexpected keyword argument ‘provider_options’” when running quick example of insightface得知：是由于onnxruntime 版本過低導(dǎo)致，更新版本

pip install onnxruntime==1.6.0

import cv2import numpy as npimport insightfacefrom insightface.app import FaceAnalysisfrom insightface.data import get_image as ins_get_imageapp = FaceAnalysis(providers=["CUDAExecutionProvider", "CPUExecutionProvider"])app.prepare(ctx_id=0, det_size=(640, 640))img = ins_get_image("C:\\Users\\supre\\Desktop\\faces\\all")faces = app.get(img)rimg = app.draw_on(img, faces)cv2.imwrite("./all_output.jpg", rimg)

五、SeetaFaceEngine

SeetaFaceEngine 是一個開源的C++人臉識別引擎，由中科院計(jì)算所山世光研究員帶領(lǐng)的人臉識別研究組研發(fā)。代碼基于C++實(shí)現(xiàn)，且不依賴于任何第三方的庫函數(shù)，開源協(xié)議為BSD-2，可供學(xué)術(shù)界和工業(yè)界免費(fèi)使用它可以運(yùn)行在CPU上。它包含人臉檢測、人臉對準(zhǔn)和人臉識別三個關(guān)鍵部分，是構(gòu)建真實(shí)人臉識別應(yīng)用系統(tǒng)的必要和充分條件

5.1 編譯

SeetaFaceEngine 包含三部分，所以需要使用 cmake 編譯三次，編譯方法見 readme

5.2 人臉檢測

#include#include#include#include#include "opencv2/highgui/highgui.hpp"#include "opencv2/imgproc/imgproc.hpp"#include "face_detection.h"using namespace std;int main(int argc, char** argv) {  const char* img_path = "C:\\Users\\supre\\Desktop\\faces\\all.jpg";  seeta::FaceDetection detector("E:\\tmp\\SeetaFaceEngine-master\\FaceDetection\\model\\seeta_fd_frontal_v1.0.bin");  detector.SetMinFaceSize(40);  detector.SetScoreThresh(2.f);  detector.SetImagePyramidScaleFactor(0.8f);  detector.SetWindowStep(4, 4);  cv::Mat img = cv::imread(img_path, cv::IMREAD_UNCHANGED);  cv::Mat img_gray;  if (img.channels() != 1)    cv::cvtColor(img, img_gray, cv::COLOR_BGR2GRAY);  else    img_gray = img;  seeta::ImageData img_data;  img_data.data = img_gray.data;  img_data.width = img_gray.cols;  img_data.height = img_gray.rows;  img_data.num_channels = 1;  long t0 = cv::getTickCount();  std::vectorfaces = detector.Detect(img_data);  long t1 = cv::getTickCount();  double secs = (t1 - t0)/cv::getTickFrequency();  cout << "Detections takes " << secs << " seconds " << endl;  cout << "Image size (wxh): " << img_data.width << "x"      << img_data.height << endl;  cv::Rect face_rect;  int32_t num_face = static_cast(faces.size());  for (int32_t i = 0; i < num_face; i++) {    face_rect.x = faces[i].bbox.x;    face_rect.y = faces[i].bbox.y;    face_rect.width = faces[i].bbox.width;    face_rect.height = faces[i].bbox.height;    cv::rectangle(img, face_rect, CV_RGB(0, 0, 255), 4, 8, 0);  }  cv::namedWindow("Test", cv::WINDOW_AUTOSIZE);  cv::imwrite("all_1.jpg", img);  cv::imshow("Test", img);  cv::waitKey(0);  cv::destroyAllWindows();}

5.3 face alignment

face alignment 指 通過一定量的訓(xùn)練集（人臉圖像和每個圖像上相對應(yīng)的多個landmarks），來得到一個model，使得該model再輸入了一張任意姿態(tài)下的人臉照片后，能夠?qū)υ撜掌械年P(guān)鍵點(diǎn)進(jìn)行標(biāo)記.

#include#include#include#include#include "cv.h"#include "highgui.h"#include "opencv2/highgui/highgui.hpp"#include "opencv2/imgproc/imgproc.hpp"#include "face_detection.h"#include "face_alignment.h"int main(int argc, char** argv){// Initialize face detection model  std::string MODEL_DIR = "E:\\tmp\\SeetaFaceEngine-master\\FaceAlignment\\model\\";  std::string DATA_DIR = "E:\\tmp\\SeetaFaceEngine-master\\FaceAlignment\\data\\";  std::string IMG_PATH = DATA_DIR + "all.jpg";  int pts_num = 5;  seeta::FaceDetection detector("E:\\tmp\\SeetaFaceEngine-master\\FaceDetection\\model\\seeta_fd_frontal_v1.0.bin");  detector.SetMinFaceSize(40);  detector.SetScoreThresh(2.f);  detector.SetImagePyramidScaleFactor(0.8f);  detector.SetWindowStep(4, 4);  // Initialize face alignment model  seeta::FaceAlignment point_detector((MODEL_DIR + "seeta_fa_v1.1.bin").c_str());  //load image  cv::Mat img = cv::imread(IMG_PATH, cv::IMREAD_UNCHANGED);  cv::Mat img_gray;  if (img.channels() != 1)    cv::cvtColor(img, img_gray, cv::COLOR_BGR2GRAY);  else    img_gray = img;  seeta::ImageData img_data;  img_data.data = img_gray.data;  img_data.width = img_gray.cols;  img_data.height = img_gray.rows;  img_data.num_channels = 1;  std::vectorfaces = detector.Detect(img_data);  int32_t face_num = static_cast(faces.size());  std::cout<<"face_num:"<<FACE_NUM; if="" (face_num="=" 0)="" {return 0;  }  cv::Rect face_rect;  for (int32_t i = 0; i < face_num; i++) {face_rect.x = faces[i].bbox.x;    face_rect.y = faces[i].bbox.y;    face_rect.width = faces[i].bbox.width;    face_rect.height = faces[i].bbox.height;    cv::rectangle(img, face_rect, CV_RGB(0, 0, 255), 4, 8, 0);    // Detect 5 facial landmarks   seeta::FacialLandmark points[5];   point_detector.PointDetectLandmarks(img_data, faces[i], points);   for (int i = 0; i< pts_num; i++)   {cv::circle(img, cvPoint(points[i].x, points[i].y), 2, CV_RGB(0, 255, 0), CV_FILLED);   }  }  cv::namedWindow("Test", cv::WINDOW_AUTOSIZE);  cv::imwrite("test.jpg", img);  cv::imshow("Test", img);  cv::waitKey(0);  cv::destroyAllWindows();  return 0;}

5.4 人臉檢測相似率

#include#include#include "opencv2/highgui/highgui.hpp"#include "opencv2/imgproc/imgproc.hpp"#include "face_identification.h"#include "recognizer.h"#include "face_detection.h"#include "face_alignment.h"#include "math_functions.h"#include#include#include#includeusing namespace seeta;using namespace std;std::string DATA_DIR = "E:\\tmp\\SeetaFaceEngine-master\\FaceIdentification\\data\\";std::string MODEL_DIR = "E:\\tmp\\SeetaFaceEngine-master\\FaceIdentification\\model\\";int main(int argc, char* argv[]) {// Initialize face detection model  seeta::FaceDetection detector("E:\\tmp\\SeetaFaceEngine-master\\FaceDetection\\model\\seeta_fd_frontal_v1.0.bin");  detector.SetMinFaceSize(40);  detector.SetScoreThresh(2.f);  detector.SetImagePyramidScaleFactor(0.8f);  detector.SetWindowStep(4, 4);  // Initialize face alignment model  seeta::FaceAlignment point_detector("E:\\tmp\\SeetaFaceEngine-master\\FaceAlignment\\model\\seeta_fa_v1.1.bin");  // Initialize face Identification model  FaceIdentification face_recognizer((MODEL_DIR + "seeta_fr_v1.0.bin").c_str());  std::string test_dir = DATA_DIR + "test_face_recognizer/";  //load image  cv::Mat gallery_img_color = cv::imread(test_dir + "images/liushishi_1.jpg", 1);  cv::Mat gallery_img_gray;  cv::cvtColor(gallery_img_color, gallery_img_gray, CV_BGR2GRAY);  cv::Mat probe_img_color = cv::imread(test_dir + "images/liushishi_2.jpg", 1);  cv::Mat probe_img_gray;  cv::cvtColor(probe_img_color, probe_img_gray, CV_BGR2GRAY);  ImageData gallery_img_data_color(gallery_img_color.cols, gallery_img_color.rows, gallery_img_color.channels());  gallery_img_data_color.data = gallery_img_color.data;  ImageData gallery_img_data_gray(gallery_img_gray.cols, gallery_img_gray.rows, gallery_img_gray.channels());  gallery_img_data_gray.data = gallery_img_gray.data;  ImageData probe_img_data_color(probe_img_color.cols, probe_img_color.rows, probe_img_color.channels());  probe_img_data_color.data = probe_img_color.data;  ImageData probe_img_data_gray(probe_img_gray.cols, probe_img_gray.rows, probe_img_gray.channels());  probe_img_data_gray.data = probe_img_gray.data;  // Detect faces  std::vectorgallery_faces = detector.Detect(gallery_img_data_gray);  int32_t gallery_face_num = static_cast(gallery_faces.size());  std::vectorprobe_faces = detector.Detect(probe_img_data_gray);  int32_t probe_face_num = static_cast(probe_faces.size());  if (gallery_face_num == 0 || probe_face_num==0)  {std::cout << "Faces are not detected.";    return 0;  }  // Detect 5 facial landmarks  seeta::FacialLandmark gallery_points[5];  point_detector.PointDetectLandmarks(gallery_img_data_gray, gallery_faces[0], gallery_points);  seeta::FacialLandmark probe_points[5];  point_detector.PointDetectLandmarks(probe_img_data_gray, probe_faces[0], probe_points);  for (int i = 0; i<5; i++)  {cv::circle(gallery_img_color, cv::Point(gallery_points[i].x, gallery_points[i].y), 2,      CV_RGB(0, 255, 0));    cv::circle(probe_img_color, cv::Point(probe_points[i].x, probe_points[i].y), 2,      CV_RGB(0, 255, 0));  }  cv::imwrite("gallery_point_result.jpg", gallery_img_color);  cv::imwrite("probe_point_result.jpg", probe_img_color);  // Extract face identity feature  float gallery_fea[2048];  float probe_fea[2048];  face_recognizer.ExtractFeatureWithCrop(gallery_img_data_color, gallery_points, gallery_fea);  face_recognizer.ExtractFeatureWithCrop(probe_img_data_color, probe_points, probe_fea);  // Caculate similarity of two faces  float sim = face_recognizer.CalcSimilarity(gallery_fea, probe_fea);  std::cout << "相似率："<<SIM <<endl;="" return="" 0;}

使用兩張劉詩詩照片對比：相似率為 0.679915

六、OpenFace

OpenFace, 一個旨在為計(jì)算機(jī)視覺和機(jī)器學(xué)習(xí)研究人員、情感計(jì)算社區(qū)和有興趣構(gòu)建基于面部行為分析的交互式應(yīng)用程序的人使用的工具。OpenFace是第一個能夠進(jìn)行面部地標(biāo)檢測、頭部姿態(tài)估計(jì)、面部動作單元識別和眼睛-注視估計(jì)的工具包，它具有可用的源代碼，可用于運(yùn)行和訓(xùn)練模型。代表 OpenFace 核心的計(jì)算機(jī)視覺算法在上述所有任務(wù)中都展示了最先進(jìn)的結(jié)果。此外，我們的工具能夠?qū)崟r性能，并能夠運(yùn)行在非專業(yè)的硬件上, 例如一個簡單的網(wǎng)絡(luò)攝像頭。

6.1 安裝

window 32位：https://github.com/TadasBaltrusaitis/OpenFace/releases/download/OpenFace_2.2.0/OpenFace_2.2.0_win_x86.zip64位：https://github.com/TadasBaltrusaitis/OpenFace/releases/download/OpenFace_2.2.0/OpenFace_2.2.0_win_x64.zip Linux：https://github.com/TadasBaltrusaitis/OpenFace/wiki/Unix-InstallationMac：https://github.com/TadasBaltrusaitis/OpenFace/wiki/Mac-Installation

6.2 運(yùn)行

OpenFace windows 版安裝完成后還需要下載模型數(shù)據(jù)：https://github.com/TadasBaltrusaitis/OpenFace/wiki/Model-download，放在安裝目錄\model\patch_experts下面。

OpenFace 還提供了一些工具用于在命令行實(shí)現(xiàn)人臉識別

FaceLandmarkImg 從照片中識別人臉，還是以仙劍3海報(bào)做例子放在samples下面，再新建輸出文件夾out_dir，開始識別人臉

FaceLandmarkImg.exe -f "samples/all.jpg" -out_dir "out_dir"

輸出結(jié)果為：

FaceLandmarkVid 從視頻中識別人臉

FaceLandmarkVidMulti 從多個視頻中識別人臉

FeatureExtraction 用于包含單個人臉的分析

參考

windows下安裝cython-bbox失敗