人脸任务在计算机视觉领域中十分重要，本项目主要使用了两类技术：人脸检测+人脸识别。
代码分为两部分内容：人脸注册 和 人脸识别
- 人脸注册：将人脸特征存储进数据库，这里用feature.csv代替
- 人脸识别：将人脸特征与CSV文件中人脸特征进行比较，如果成功匹配则写入考勤文件attendance.csv
文章前半部分为一步步实现流程介绍，最后会有整理过后的完整项目代码。
一、项目实现
A. 注册：
导入相关包
import cv2 import numpy as np import dlib import time import csv # from argparse import ArgumentParser from PIL import Image, ImageDraw, ImageFont
设计注册功能
注册过程我们需要完成的事：
- 打开摄像头获取画面图片
- 在图片中检测并获取人脸位置
- 根据人脸位置获取68个关键点
- 根据68个关键点生成特征描述符
- 保存
- (优化)展示界面，加入注册时成功提示等
1、基本步骤
我们首先进行前三步：
# 检测人脸，获取68个关键点，获取特征描述符 def faceRegister(faceId=1, userName='default', interval=3, faceCount=3, resize_w=700, resize_h=400): ''' faceId:人脸ID userName: 人脸姓名 faceCount: 采集该人脸图片的数量 interval: 采集间隔 ''' cap = cv2.VideoCapture(0) # 人脸检测模型 hog_face_detector = dlib.get_frontal_face_detector() # 关键点 检测模型 shape_detector = dlib.shape_predictor('./weights/shape_predictor_68_face_landmarks.dat') # resnet模型 face_descriptor_extractor = dlib.face_recognition_model_v1('./weights/dlib_face_recognition_resnet_model_v1.dat') while True: ret, frame = cap.read() # 镜像 frame = cv2.flip(frame,1) # 转为灰度图 frame_gray = cv2.cvtColor(frame,cv2.COLOR_BGR2GRAY) # 检测人脸 detections = hog_face_detector(frame,1) for face in detections: # 人脸框坐标 左上和右下 l, t, r, b = face.left(), face.top(), face.right(), face.bottom() # 获取68个关键点 points = shape_detector(frame,face) # 绘制关键点 for point in points.parts(): cv2.circle(frame,(point.x,point.y),2,(0,255,0),1) # 绘制矩形框 cv2.rectangle(frame,(l,t),(r,b),(0,255,0),2) cv2.imshow("face",frame) if cv2.waitKey(10) & 0xFF == ord('q'): break cap.release() cv2.destroyAllWindows faceRegister()此时一张帅脸如下：
2、描述符的采集
之后，我们根据参数，即faceCount 和 Interval 进行描述符的生成和采集。
(这里我默认是faceCount=3，Interval=3，即每3秒采集一次，共3次）
def faceRegister(faceId=1, userName='default', interval=3, faceCount=3, resize_w=700, resize_h=400): ''' faceId:人脸ID userName: 人脸姓名 faceCount: 采集该人脸图片的数量 interval: 采集间隔 ''' cap = cv2.VideoCapture(0) # 人脸检测模型 hog_face_detector = dlib.get_frontal_face_detector() # 关键点 检测模型 shape_detector = dlib.shape_predictor('./weights/shape_predictor_68_face_landmarks.dat') # resnet模型 face_descriptor_extractor = dlib.face_recognition_model_v1('./weights/dlib_face_recognition_resnet_model_v1.dat') # 开始时间 start_time = time.time() # 执行次数 collect_times = 0 while True: ret, frame = cap.read() # 镜像 frame = cv2.flip(frame,1) # 转为灰度图 frame_gray = cv2.cvtColor(frame,cv2.COLOR_BGR2GRAY) # 检测人脸 detections = hog_face_detector(frame,1) for face in detections: # 人脸框坐标 左上和右下 l, t, r, b = face.left(), face.top(), face.right(), face.bottom() # 获取68个关键点 points = shape_detector(frame,face) # 绘制人脸关键点 for point in points.parts(): cv2.circle(frame, (point.x, point.y), 2, (0, 255, 0), 1) # 绘制矩形框 cv2.rectangle(frame, (l, t), (r, b), (0, 255, 0), 2) # 采集： if collect_times interval: # 获取特征描述符 face_descriptor = face_descriptor_extractor.compute_face_descriptor(frame,points) # dlib格式转为数组 face_descriptor = [f for f in face_descriptor] collect_times += 1 start_time = now print("成功采集{}次".format(collect_times)) else: # 时间间隔不到interval print("等待进行下一次采集") pass else: # 已经成功采集完3次了 print("采集完毕") cap.release() cv2.destroyAllWindows() return cv2.imshow("face",frame) if cv2.waitKey(10) & 0xFF == ord('q'): break cap.release() cv2.destroyAllWindows() faceRegister()等待进行下一次采集 ... 成功采集1次 等待进行下一次采集 ... 成功采集2次 等待进行下一次采集 ... 成功采集3次 采集完毕
3、完整的注册
最后就是写入csv文件
这里加入了注册成功等的提示，且把一些变量放到了全局，因为后面人脸识别打卡时也会用到。
# 加载人脸检测器 hog_face_detector = dlib.get_frontal_face_detector() cnn_detector = dlib.cnn_face_detection_model_v1('./weights/mmod_human_face_detector.dat') haar_face_detector = cv2.CascadeClassifier('./weights/haarcascade_frontalface_default.xml') # 加载关键点检测器 points_detector = dlib.shape_predictor('./weights/shape_predictor_68_face_landmarks.dat') # 加载resnet模型 face_descriptor_extractor = dlib.face_recognition_model_v1('./weights/dlib_face_recognition_resnet_model_v1.dat')# 绘制中文 def cv2AddChineseText(img, text, position, textColor=(0, 255, 0), textSize=30): if (isinstance(img, np.ndarray)): # 判断是否OpenCV图片类型 img = Image.fromarray(cv2.cvtColor(img, cv2.COLOR_BGR2RGB)) # 创建一个可以在给定图像上绘图的对象 draw = ImageDraw.Draw(img) # 字体的格式 fontStyle = ImageFont.truetype( "./fonts/songti.ttc", textSize, encoding="utf-8") # 绘制文本 draw.text(position, text, textColor, font=fontStyle) # 转换回OpenCV格式 return cv2.cvtColor(np.asarray(img), cv2.COLOR_RGB2BGR)# 绘制左侧信息 def drawLeftInfo(frame, fpsText, mode="Reg", detector='haar', person=1, count=1): # 帧率 cv2.putText(frame, "FPS: " + str(round(fpsText, 2)), (30, 50), cv2.FONT_ITALIC, 0.8, (0, 255, 0), 2) # 模式：注册、识别 cv2.putText(frame, "Mode: " + str(mode), (30, 80), cv2.FONT_ITALIC, 0.8, (0, 255, 0), 2) if mode == 'Recog': # 检测器 cv2.putText(frame, "Detector: " + detector, (30, 110), cv2.FONT_ITALIC, 0.8, (0, 255, 0), 2) # 人数 cv2.putText(frame, "Person: " + str(person), (30, 140), cv2.FONT_ITALIC, 0.8, (0, 255, 0), 2) # 总人数 cv2.putText(frame, "Count: " + str(count), (30, 170), cv2.FONT_ITALIC, 0.8, (0, 255, 0), 2)# 注册人脸 def faceRegiser(faceId=1, userName='default', interval=3, faceCount=3, resize_w=700, resize_h=400): # 计数 count = 0 # 开始注册时间 startTime = time.time() # 视频时间 frameTime = startTime # 控制显示打卡成功的时长 show_time = (startTime - 10) # 打开文件 f = open('./data/feature.csv', 'a', newline='') csv_writer = csv.writer(f) cap = cv2.VideoCapture(0) while True: ret, frame = cap.read() frame = cv2.resize(frame, (resize_w, resize_h)) frame = cv2.flip(frame, 1) # 检测 face_detetion = hog_face_detector(frame, 1) for face in face_detetion: # 识别68个关键点 points = points_detector(frame, face) # 绘制人脸关键点 for point in points.parts(): cv2.circle(frame, (point.x, point.y), 2, (0, 255, 0), 1) # 绘制框框 l, t, r, b = face.left(), face.top(), face.right(), face.bottom() cv2.rectangle(frame, (l, t), (r, b), (0, 255, 0), 2) now = time.time() if (now - show_time) interval: # 特征描述符 face_descriptor = face_descriptor_extractor.compute_face_descriptor(frame, points) face_descriptor = [f for f in face_descriptor] # 描述符增加进data文件 line = [faceId, userName, face_descriptor] # 写入 csv_writer.writerow(line) # 保存照片样本 print('人脸注册成功 {count}/{faceCount}，faceId:{faceId}，userName:{userName}'.format(count=(count + 1), faceCount=faceCount, faceId=faceId, userName=userName)) frame = cv2AddChineseText(frame, "注册成功 {count}/{faceCount}".format(count=(count + 1), faceCount=faceCount), (l, b + 30), textColor=(255, 0, 255), textSize=30) show_time = time.time() # 时间重置 startTime = now # 次数加一 count += 1 else: print('人脸注册完毕') f.close() cap.release() cv2.destroyAllWindows() return now = time.time() fpsText = 1 / (now - frameTime) frameTime = now # 绘制 drawLeftInfo(frame, fpsText, 'Register') cv2.imshow('Face Attendance Demo: Register', frame) if cv2.waitKey(10) & 0xFF == ord('q'): break f.close() cap.release() cv2.destroyAllWindows()此时执行：
faceRegiser(3,"用户B")
人脸注册成功 1/3，faceId:3，userName:用户B 人脸注册成功 2/3，faceId:3，userName:用户B 人脸注册成功 3/3，faceId:3，userName:用户B 人脸注册完毕
其features文件：
B. 识别、打卡
识别步骤如下：
- 打开摄像头获取画面
- 根据画面中的图片获取里面的人脸特征描述符
- 根据特征描述符将其与feature.csv文件里特征做距离判断
- 获取ID、NAME
- 考勤记录写入attendance.csv里
这里与上面流程相似，不过是加了一个对比功能，距离小于阈值，则表示匹配成功。就加快速度不一步步来了，代码如下：
# 刷新右侧考勤信息 def updateRightInfo(frame, face_info_list, face_img_list): # 重新绘制逻辑：从列表中每隔3个取一批显示，新增人脸放在最前面 # 如果有更新，重新绘制 # 如果没有，定时往后移动 left_x = 30 left_y = 20 resize_w = 80 offset_y = 120 index = 0 frame_h = frame.shape[0] frame_w = frame.shape[1] for face in face_info_list[:3]: name = face[0] time = face[1] face_img = face_img_list[index] # print(face_img.shape) face_img = cv2.resize(face_img, (resize_w, resize_w)) offset_y_value = offset_y * index frame[(left_y + offset_y_value):(left_y + resize_w + offset_y_value), -(left_x + resize_w):-left_x] = face_img cv2.putText(frame, name, ((frame_w - (left_x + resize_w)), (left_y + resize_w) + 15 + offset_y_value), cv2.FONT_ITALIC, 0.5, (0, 255, 0), 1) cv2.putText(frame, time, ((frame_w - (left_x + resize_w)), (left_y + resize_w) + 30 + offset_y_value), cv2.FONT_ITALIC, 0.5, (0, 255, 0), 1) index += 1 return frame# 返回DLIB格式的face def getDlibRect(detector='hog', face=None): l, t, r, b = None, None, None, None if detector == 'hog': l, t, r, b = face.left(), face.top(), face.right(), face.bottom() if detector == 'cnn': l = face.rect.left() t = face.rect.top() r = face.rect.right() b = face.rect.bottom() if detector == 'haar': l = face[0] t = face[1] r = face[0] + face[2] b = face[1] + face[3] nonnegative = lambda x: x if x >= 0 else 0 return map(nonnegative, (l, t, r, b))
# 获取CSV中信息 def getFeatList(): print('加载注册的人脸特征') feature_list = None label_list = [] name_list = [] # 加载保存的特征样本 with open('./data/feature.csv', 'r') as f: csv_reader = csv.reader(f) for line in csv_reader: # 重新加载数据 faceId = line[0] userName = line[1] face_descriptor = eval(line[2]) label_list.append(faceId) name_list.append(userName) # 转为numpy格式 face_descriptor = np.asarray(face_descriptor, dtype=np.float64) # 转为二维矩阵，拼接 face_descriptor = np.reshape(face_descriptor, (1, -1)) # 初始化 if feature_list is None: feature_list = face_descriptor else: # 拼接 feature_list = np.concatenate((feature_list, face_descriptor), axis=0) print("特征加载完毕") return feature_list, label_list, name_list# 人脸识别 def faceRecognize(detector='haar', threshold=0.5, write_video=False, resize_w=700, resize_h=400): # 视频时间 frameTime = time.time() # 加载特征 feature_list, label_list, name_list = getFeatList() face_time_dict = {} # 保存name,time人脸信息 face_info_list = [] # numpy格式人脸图像数据 face_img_list = [] # 侦测人数 person_detect = 0 # 统计人脸数 face_count = 0 # 控制显示打卡成功的时长 show_time = (frameTime - 10) # 考勤记录 f = open('./data/attendance.csv', 'a') csv_writer = csv.writer(f) cap = cv2.VideoCapture(0) # resize_w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))//2 # resize_h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT)) //2 videoWriter = cv2.VideoWriter('./record_video/out' + str(time.time()) + '.mp4', cv2.VideoWriter_fourcc(*'MP4V'), 15, (resize_w, resize_h)) while True: ret, frame = cap.read() frame = cv2.resize(frame, (resize_w, resize_h)) frame = cv2.flip(frame, 1) # 切换人脸检测器 if detector == 'hog': face_detetion = hog_face_detector(frame, 1) if detector == 'cnn': face_detetion = cnn_detector(frame, 1) if detector == 'haar': frame_gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY) face_detetion = haar_face_detector.detectMultiScale(frame_gray, minNeighbors=7, minSize=(100, 100)) person_detect = len(face_detetion) for face in face_detetion: l, t, r, b = getDlibRect(detector, face) face = dlib.rectangle(l, t, r, b) # 识别68个关键点 points = points_detector(frame, face) cv2.rectangle(frame, (l, t), (r, b), (0, 255, 0), 2) # 人脸区域 face_crop = frame[t:b, l:r] # 特征 face_descriptor = face_descriptor_extractor.compute_face_descriptor(frame, points) face_descriptor = [f for f in face_descriptor] face_descriptor = np.asarray(face_descriptor, dtype=np.float64) # 计算距离 distance = np.linalg.norm((face_descriptor - feature_list), axis=1) # 最小距离索引 min_index = np.argmin(distance) # 最小距离 min_distance = distance[min_index] predict_name = "Not recog" if min_distance 3秒，或者换了一个人，将这条记录插入 need_insert = False now = time.time() if predict_name in face_time_dict: if (now - face_time_dict[predict_name]) > 3: # 刷新时间 face_time_dict[predict_name] = now need_insert = True else: # 还是上次人脸 need_insert = False else: # 新增数据记录 face_time_dict[predict_name] = now need_insert = True if (now - show_time) 10: face_info_list = face_info_list[:9] face_img_list = face_img_list[:9] frame = updateRightInfo(frame, face_info_list, face_img_list) if write_video: videoWriter.write(frame) cv2.imshow('Face Attendance Demo: Recognition', frame) if cv2.waitKey(10) & 0xFF == ord('q'): break f.close() videoWriter.release() cap.release() cv2.destroyAllWindows()然后效果就和我们宿舍楼下差不多了~
我年轻的时候，我大概比现在帅个几百倍吧，哎。
二、总代码
上文其实把登录和注册最后一部分代码放在一起就是了，这里就不再复制粘贴了，相关权重文件下载链接：opencv/data at master · opencv/opencv · GitHub
当然本项目还有很多需要优化的地方，比如设置用户不能重复、考勤打卡每天只能一次、把csv改为链接成数据库等等，后续代码优化完成后就可以部署然后和室友**了。
