您好,登錄后才能下訂單哦!
最近開(kāi)始學(xué)習(xí)Qt,結(jié)合之前學(xué)習(xí)過(guò)的caffe一起搭建了一個(gè)人臉識(shí)別登錄系統(tǒng)的程序,新手可能有理解不到位的情況,還請(qǐng)大家多多指教。
我的想法是用opencv自帶的人臉檢測(cè)算法檢測(cè)出面部,利用caffe訓(xùn)練好的卷積神經(jīng)網(wǎng)絡(luò)來(lái)提取特征,通過(guò)計(jì)算當(dāng)前檢測(cè)到的人臉與已近注冊(cè)的所有用戶的面部特征之間的相似度,如果最大的相似度大于一個(gè)閾值,就可以確定當(dāng)前檢測(cè)到的人臉對(duì)應(yīng)為這個(gè)相似度最大的用戶了。
###Caffe人臉識(shí)別
因?yàn)椴粩嘤行碌挠脩艏尤耄欢砑有掠脩艉笾匦抡{(diào)整CNN的網(wǎng)絡(luò)結(jié)構(gòu)太費(fèi)時(shí)間,所以不能用CNN去判別一個(gè)用戶屬于哪一類。一個(gè)訓(xùn)練好的人臉識(shí)別網(wǎng)絡(luò)擁有很強(qiáng)大的特征提取能力(例如這里用到的VGG face),我們finetune預(yù)訓(xùn)練的網(wǎng)絡(luò)時(shí)會(huì)調(diào)整最后一層的分類數(shù)目,所以最后一層的目的是為了分類,倒數(shù)第二個(gè)全連接層(或者前面的)提取到的特征通過(guò)簡(jiǎn)單的計(jì)算距離也可以達(dá)到很高的準(zhǔn)確率,因此可以用計(jì)算相似度的方式判斷類別。
載入finetune后的VGG模型
代碼就不詳細(xì)解釋了,我用的是拿1000個(gè)人臉微調(diào)后的VGGface,效果比用直接下載來(lái)的weight文件好一點(diǎn),這里可以用原始的權(quán)重文件代替。
import caffe model_def = 'VGG_FACE_deploy.prototxt' model_weights = 'VGG_Face_finetune_1000_iter_900.caffemodel' # create transformer for the input called 'data' net = caffe.Net(model_def, # defines the structure of the model model_weights, # contains the trained weights caffe.TEST) transformer = caffe.io.Transformer({'data': net.blobs['data'].data.shape}) transformer.set_transpose('data', (2,0,1)) # move image channels to outermost dimension transformer.set_mean('data', np.array([104, 117, 123])) # subtract the dataset-mean value in each channel transformer.set_raw_scale('data', 255) # rescale from [0, 1] to [0, 255] transformer.set_channel_swap('data', (2,1,0)) # swap channels from RGB to BGRxpor
計(jì)算余弦相似度
import numpy as np # 計(jì)算余弦距離 def cal_cos(A,B): num = A.dot(B.T) #若為行向量則 A * B.T print(B.shape) if B.ndim == 1: denom = np.linalg.norm(A) * np.linalg.norm(B) else: denom = np.linalg.norm(A) * np.linalg.norm(B, axis=1) #print(num) cos = num / denom #余弦值 sim = 0.5 + 0.5 * cos #歸一化 return sim def cal_feature(image): #for i,img_name in enumerate(os.listdir(path)): #image = caffe.io.load_image(os.path.join(path,img_name)) transformed_image = transformer.preprocess('data', image) net.blobs['data'].data[0,:,:,:] = transformed_image output = net.forward() return net.blobs['fc7'].data[0]
cal_feature函數(shù)返回fc7層的輸出,也就是image通過(guò)網(wǎng)絡(luò)提取到的特征;A的維度為[1, 4096],為需要檢測(cè)的目標(biāo),B的維度為[n,4096],表示所有已注冊(cè)的用戶的特征,cal_cos返回n個(gè)相似度,值越大,越可能是同一個(gè)人。
###Opencv人臉檢測(cè)
檢測(cè)人臉位置的算法用了opencv自帶的人臉檢測(cè)器。
import cv2 face_cascade = cv2.CascadeClassifier('haarcascade_frontalface_default.xml')
PyQt界面
定義全局變量存儲(chǔ)用戶的信息,提取到的特征,我用文件的形式將這些信息保存到本地,下一次運(yùn)行時(shí)提前載入。
import sys import os import pickle global ALLFEATURE, NEWFEATURE, tempUsrName, ALLUSER, USRNAME with open('USRNAME.pickle', 'rb') as f: USRNAME = pickle.load(f) with open('ALLUSER.pickle', 'rb') as f: ALLUSER = pickle.load(f) ALLFEATURE = np.load('usrfeature.npy') NEWFEATURE = np.array([]) tempUsrName = {}
設(shè)計(jì)一個(gè)登錄界面
用PyQt設(shè)計(jì)一個(gè)界面,實(shí)現(xiàn)用戶注冊(cè),注冊(cè)時(shí)錄入照片,用戶密碼登錄,人臉識(shí)別登錄等功能。
創(chuàng)建一個(gè)TabWidget界面
tab1用來(lái)實(shí)現(xiàn)密碼登錄,注冊(cè),tab2用來(lái)實(shí)現(xiàn)人臉識(shí)別登錄。
from PyQt5.QtWidgets import (QWidget, QMessageBox, QLabel, QDialog, QApplication, QPushButton, QDesktopWidget, QLineEdit, QTabWidget) from PyQt5.QtGui import QIcon, QPixmap, QImage, QPalette, QBrush from PyQt5.QtCore import Qt, QTimer class TabWidget(QTabWidget): def __init__(self, parent=None): super(TabWidget, self).__init__(parent) self.setWindowTitle('Face Recognition') self.setWindowIcon(QIcon('camera.png')) self.resize(400, 260) self.center() self.mContent = passWordSign() self.mIndex = faceSign() self.addTab(self.mContent, QIcon('camera.png'), u"密碼登錄") self.addTab(self.mIndex, u"人臉識(shí)別") palette=QPalette() icon=QPixmap('background.jpg').scaled(400, 260) palette.setBrush(self.backgroundRole(), QBrush(icon)) #添加背景圖片 self.setPalette(palette) def center(self): qr = self.frameGeometry() cp = QDesktopWidget().availableGeometry().center() qr.moveCenter(cp) self.move(qr.topLeft()) def closeEvent(self, event): reply = QMessageBox.question(self, 'Message', "Are you sure to quit?", QMessageBox.Yes | QMessageBox.No, QMessageBox.No) if reply == QMessageBox.Yes: event.accept() else: event.ignore() if __name__ == '__main__': app = QApplication(sys.argv) t = TabWidget() t.show() #ex = Example() sys.exit(app.exec_())
用戶注冊(cè)和密碼登錄
分別添加兩個(gè)按鈕和兩個(gè)文本框,文本框用于用戶名和密碼輸入,按鈕分別對(duì)應(yīng)事件注冊(cè)和登錄。addPicture用于注冊(cè)時(shí)錄入用戶照片。
class passWordSign(QWidget): def __init__(self): super().__init__() self.initUI() def initUI(self): #self.setGeometry(0, 0, 450, 300) self.signUpButton = QPushButton(QIcon('camera.png'), 'Sign up', self) self.signUpButton.move(300, 200) self.signInButton = QPushButton(QIcon('camera.png'), 'Sign in', self) self.signInButton.move(200, 200) self.usrNameLine = QLineEdit( self ) self.usrNameLine.setPlaceholderText('User Name') self.usrNameLine.setFixedSize(200, 30) self.usrNameLine.move(100, 50) self.passWordLine = QLineEdit(self) self.passWordLine.setEchoMode(QLineEdit.Password) self.passWordLine.setPlaceholderText('Pass Word') self.passWordLine.setFixedSize(200, 30) self.passWordLine.move(100, 120) self.signInButton.clicked.connect(self.signIn) self.signUpButton.clicked.connect(self.signUp) self.show() def signIn(self): global ALLFEATURE, NEWFEATURE, tempUsrName, ALLUSER, USRNAME if self.usrNameLine.text() not in ALLUSER: QMessageBox.information(self,"Information","用戶不存在,請(qǐng)注冊(cè)") elif ALLUSER[self.usrNameLine.text()] == self.passWordLine.text(): QMessageBox.information(self,"Information","Welcome!") else: QMessageBox.information(self,"Information","密碼錯(cuò)誤!") def signUp(self): global ALLFEATURE, NEWFEATURE, tempUsrName, ALLUSER, USRNAME if self.usrNameLine.text() in ALLUSER: QMessageBox.information(self,"Information","用戶已存在!") elif len(self.passWordLine.text()) < 3: QMessageBox.information(self,"Information","密碼太短!") else: tempUsrName.clear() tempUsrName[self.usrNameLine.text()] = self.passWordLine.text() self.addPicture() def addPicture(self): dialog = Dialog(parent=self) dialog.show()
錄入用戶人臉
點(diǎn)擊sign up按鈕后彈出一個(gè)對(duì)話框,用一個(gè)label控件顯示攝像頭獲取的照片。首先用opencv打開(kāi)攝像頭,用自帶的人臉檢測(cè)器檢測(cè)到人臉self.face后,繪制一個(gè)藍(lán)色的框,然后resize到固定的大?。▽?duì)應(yīng)網(wǎng)絡(luò)的輸入)。將opencv的圖片格式轉(zhuǎn)換為Qlabel可以顯示的格式,用Qtimer定時(shí)器每隔一段時(shí)間刷新圖片。檢測(cè)鼠標(biāo)點(diǎn)擊事件mousePressEvent,點(diǎn)擊鼠標(biāo)后保存當(dāng)前錄入的用戶注冊(cè)信息和對(duì)應(yīng)的特征。關(guān)閉攝像頭,提示注冊(cè)成功。
class Dialog(QDialog): def __init__(self, parent=None): QDialog.__init__(self, parent) self.resize(240, 200) self.label = QLabel(self) self.label.setFixedWidth(150) self.label.setFixedHeight(150) self.label.move(40, 20) pixMap = QPixmap("face.jpg").scaled(self.label.width(),self.label.height()) self.label.setPixmap(pixMap) self.label.show() self.timer = QTimer() self.timer.start() self.timer.setInterval(100) self.cap = cv2.VideoCapture(0) self.timer.timeout.connect(self.capPicture) def mousePressEvent(self, event): global ALLFEATURE, NEWFEATURE, tempUsrName, ALLUSER, USRNAME self.cap.release() NEWFEATURE = cal_feature(self.face).reshape([1,-1]) if NEWFEATURE.size > 0: for key, value in tempUsrName.items(): ALLUSER[key] = value USRNAME.append(key) with open('ALLUSER.pickle', 'wb') as f: pickle.dump(ALLUSER, f) with open('USRNAME.pickle', 'wb') as f: pickle.dump(USRNAME, f) print(ALLFEATURE,NEWFEATURE) ALLFEATURE = np.concatenate((ALLFEATURE, NEWFEATURE), axis=0) np.save('usrfeature.npy', ALLFEATURE) QMessageBox.information(self,"Information","Success!") def capPicture(self): if (self.cap.isOpened()): # get a frame ret, img = self.cap.read() gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) faces = face_cascade.detectMultiScale(gray, 1.3, 5) for (x,y,w,h) in faces: img = cv2.rectangle(img,(x,y),(x+w,y+h),(255,0,0),2) roi_gray = gray[y:y+h, x:x+w] roi_color = img[y:y+h, x:x+w] self.face = cv2.resize(img[y:y+h, x:x+w],(224, 224), interpolation=cv2.INTER_CUBIC) height, width, bytesPerComponent = img.shape bytesPerLine = bytesPerComponent * width # 變換彩色空間順序 cv2.cvtColor(img, cv2.COLOR_BGR2RGB, img) # 轉(zhuǎn)為QImage對(duì)象 self.image = QImage(img.data, width, height, bytesPerLine, QImage.Format_RGB888) self.label.setPixmap(QPixmap.fromImage(self.image).scaled(self.label.width(),self.label.height()))
人臉識(shí)別登錄
登錄部分與之前類似,添加一個(gè)label控件用來(lái)顯示圖片,兩個(gè)按鈕用來(lái)開(kāi)始檢測(cè)和選定圖片。當(dāng)最大的相似度大于0.9時(shí),顯示登錄成功。
class faceSign(QWidget): def __init__(self): super().__init__() self.initUI() def initUI(self): self.label = QLabel(self) self.label.setFixedWidth(260) self.label.setFixedHeight(200) self.label.move(20, 15) self.pixMap = QPixmap("face.jpg").scaled(self.label.width(),self.label.height()) self.label.setPixmap(self.pixMap) self.label.show() self.startButton = QPushButton('start', self) self.startButton.move(300, 50) self.capPictureButton = QPushButton('capPicture', self) self.capPictureButton.move(300, 150) self.startButton.clicked.connect(self.start) self.capPictureButton.clicked.connect(self.cap) #self.cap = cv2.VideoCapture(0) #self.ret, self.img = self.cap.read() self.timer = QTimer() self.timer.start() self.timer.setInterval(100) def start(self,event): self.cap = cv2.VideoCapture(0) self.timer.timeout.connect(self.capPicture) def cap(self,event): global ALLFEATURE, NEWFEATURE, tempUsrName, ALLUSER, USRNAME self.cap.release() feature = cal_feature(self.face) #np.save('usrfeature.npy', ALLFEATURE) sim = cal_cos(feature,np.array(ALLFEATURE)) m = np.argmax(sim) if max(sim)>0.9: print(sim, USRNAME) QMessageBox.information(self,"Information","Welcome," + USRNAME[m]) else: QMessageBox.information(self,"Information","識(shí)別失敗!") self.label.setPixmap(self.pixMap) def capPicture(self): if (self.cap.isOpened()): # get a frame ret, img = self.cap.read() gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) faces = face_cascade.detectMultiScale(gray, 1.3, 5) for (x,y,w,h) in faces: img = cv2.rectangle(img,(x,y),(x+w,y+h),(255,0,0),2) roi_gray = gray[y:y+h, x:x+w] roi_color = img[y:y+h, x:x+w] self.face = cv2.resize(img[y:y+h, x:x+w],(224, 224), interpolation=cv2.INTER_CUBIC) height, width, bytesPerComponent = img.shape bytesPerLine = bytesPerComponent * width # 變換彩色空間順序 cv2.cvtColor(img, cv2.COLOR_BGR2RGB, img) # 轉(zhuǎn)為QImage對(duì)象 self.image = QImage(img.data, width, height, bytesPerLine, QImage.Format_RGB888) self.label.setPixmap(QPixmap.fromImage(self.image).scaled(self.label.width(),self.label.height()))
###效果
密碼登錄,輸入合法的密碼后點(diǎn)擊sign in,顯示歡迎。
注冊(cè)界面
識(shí)別界面
登錄成功
點(diǎn)擊capPicture按鈕后,開(kāi)始計(jì)算相似度,大于0.9提示登錄成功,并顯示用戶名。
###缺點(diǎn)和不足
程序用pyinstaller打包后,親測(cè)可以在別的linux電腦下運(yùn)行。代碼和文件可以參考我的Github(沒(méi)有VGG face的權(quán)重),第一次寫(xiě)博客,非常感謝大家的意見(jiàn)??偨Y(jié)一下不足:
1.初始話caffe模型很費(fèi)時(shí)間,所以程序打開(kāi)時(shí)要等一兩秒;
2.用戶信息用文件的形式保存并不安全,可以用mysql保存到數(shù)據(jù)庫(kù),需要時(shí)調(diào)出;
3.人臉位置檢測(cè)可以用faster rcnn代替,再加上對(duì)齊;
4.識(shí)別很耗費(fèi)時(shí)間,因此不能實(shí)時(shí)檢測(cè),應(yīng)該可以用多線程解決。
以上就是本文的全部?jī)?nèi)容,希望對(duì)大家的學(xué)習(xí)有所幫助,也希望大家多多支持億速云。
免責(zé)聲明:本站發(fā)布的內(nèi)容(圖片、視頻和文字)以原創(chuàng)、轉(zhuǎn)載和分享為主,文章觀點(diǎn)不代表本網(wǎng)站立場(chǎng),如果涉及侵權(quán)請(qǐng)聯(lián)系站長(zhǎng)郵箱:is@yisu.com進(jìn)行舉報(bào),并提供相關(guān)證據(jù),一經(jīng)查實(shí),將立刻刪除涉嫌侵權(quán)內(nèi)容。