這個(gè)項(xiàng)目在之前人工智能課設(shè)上做過，但是當(dāng)時(shí)是劃水用的別人的。最近自己實(shí)現(xiàn)了一下，基本功能可以實(shí)現(xiàn)，但是效果并不是很好。容易出現(xiàn)錯(cuò)誤識(shí)別，或者更改了背景之后識(shí)別效果變差的現(xiàn)象。個(gè)人以為是數(shù)據(jù)選取的問題，希望路過的大佬批評指正。

所需要的庫

tensorflow-gpu 2.0.0 Keras 2.3.1 opencv-python 4.5.1.48 numpy 1.20.2

1.數(shù)據(jù)集獲取

這個(gè)數(shù)據(jù)集是通過自己電腦的攝像頭獲取的，也可以通過一段視頻獲取。這一過程是建立自己的人臉庫。通過Opencv中的CascadeClassifier人臉識(shí)別分類器來實(shí)現(xiàn)。
部分代碼：

def CatchPICFromVideo(window_name, camera_idx, catch_pic_num, path_name):cv2.namedWindow(window_name)# 視頻來源，可以來自一段已存好的視頻，也可以直接來自USB攝像頭cap = cv2.VideoCapture(camera_idx,cv2.CAP_DSHOW)# 告訴OpenCV使用人臉識(shí)別分類器classfier = cv2.CascadeClassifier('E:\ProgramData\Anaconda3\Lib\site-packages\cv2\data\haarcascade_frontalface_alt.xml')# 識(shí)別出人臉后要畫的邊框的顏色，RGB格式color = (0, 255, 0)num = 0while cap.isOpened():ok, frame = cap.read() # 讀取一幀數(shù)據(jù)if not ok:breakgrey = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY) # 將當(dāng)前楨圖像轉(zhuǎn)換成灰度圖像# 人臉檢測，1.2和2分別為圖片縮放比例和需要檢測的有效點(diǎn)數(shù)faceRects = classfier.detectMultiScale(grey, scaleFactor=1.2, minNeighbors=3, minSize=(32, 32))if len(faceRects) > 0: # 大于0則檢測到人臉for faceRect in faceRects: # 單獨(dú)框出每一張人臉x, y, w, h = faceRect# 將當(dāng)前幀保存為圖片img_name = '%s/%d.jpg' % (path_name, num)image = frame[y - 10: y + h + 10, x - 10: x + w + 10]cv2.imwrite(img_name, image)num += 1if num > (catch_pic_num): # 如果超過指定最大保存數(shù)量退出循環(huán)break# 畫出矩形框cv2.rectangle(frame, (x - 10, y - 10), (x + w + 10, y + h + 10), color, 2)# 顯示當(dāng)前捕捉到了多少人臉圖片了，這樣站在那里被拍攝時(shí)心里有個(gè)數(shù)，不用兩眼一抹黑傻等著font = cv2.FONT_HERSHEY_SIMPLEXcv2.putText(frame, 'num:%d' % (num), (x + 30, y + 30), font, 1, (255, 0, 255), 4)# 超過指定最大保存數(shù)量結(jié)束程序if num > (catch_pic_num): break# 顯示圖像cv2.imshow(window_name, frame)c = cv2.waitKey(10)if c & 0xFF == ord('q'):break# 釋放攝像頭并銷毀所有窗口cap.release()cv2.destroyAllWindows()

效果如下所示:

本人的人臉庫中一共有三個(gè)數(shù)據(jù)集：

2.數(shù)據(jù)集處理

數(shù)據(jù)集處理包括捕捉人臉+數(shù)據(jù)集灰度化，將源文件路徑中的數(shù)據(jù)經(jīng)過處理之后保存到目標(biāo)文件中。
部分代碼：

try:#讀取照片，第一個(gè)元素是文件名resultArray = readALLImg(sourcePath,*suffix)#對list中的圖片逐一進(jìn)行檢查，找出其中的人臉然后寫到目標(biāo)文件夾中count = 1face_cascade = cv2.CascadeClassifier('E:\ProgramData\Anaconda3\Lib\site-packages\cv2\data\haarcascade_frontalface_alt.xml')for i in resultArray:if type(i) != str:gray = cv2.cvtColor(i,cv2.COLOR_BGR2GRAY)faces = face_cascade.detectMultiScale(gray,1.3,5)for (x,y,w,h) in faces:listStr = [str(int(time.time())),str(count)] #以時(shí)間戳和讀取的排序作為文件名稱fileName = ''.join(listStr)f = cv2.resize(gray[y:(y+h),x:(x+w)],(200,200))cv2.imwrite(objectPath+os.sep+'%s.jpg'%fileName,f)count+=1except IOError:print('error')else:print('Already read ' + str(count-1) +' Faces to Destination '+objectPath)

經(jīng)過處理之后的效果如下所示:

注：并不是所有的圖片都會(huì)轉(zhuǎn)化為灰度圖片，因?yàn)椴蹲狡髟谧R(shí)別的時(shí)候有可能出錯(cuò)，個(gè)別的圖片無法識(shí)別。我的數(shù)據(jù)集采用了1500張圖片。

3.劃分訓(xùn)練集和測試集

訓(xùn)練集和測試集采用8：2的比例，所有的圖片經(jīng)過歸一化處理，大小為128*128的灰度圖片，隨機(jī)打亂。
部分代碼：

#將數(shù)據(jù)集打亂隨機(jī)分組X_train,X_test,y_train,y_test = train_test_split(imgs,labels,test_size=0.2,random_state=random.randint(0, 100))#重新格式化和標(biāo)準(zhǔn)化X_train = X_train.reshape(X_train.shape[0],1,self.img_size,self.img_size)/255.0X_test = X_test.reshape(X_test.shape[0],1,self.img_size,self.img_size)/255.0# 將labels轉(zhuǎn)成 binary class matricesy_train = np_utils.to_categorical(y_train, num_classes=counter)y_test = np_utils.to_categorical(y_test, num_classes=counter)

4.模型搭建

CNN模型為三層卷積池化層+兩層全連接層，抹平之后進(jìn)行分類。中間加入Dropout層，防止過擬合。
部分代碼：

def build_model(self):self.model = keras.Sequential([keras.layers.Conv2D(filters=32,kernel_size=(5,5),padding="same",activation="relu",input_shape=self.dataset.X_train.shape[1:]),keras.layers.MaxPooling2D(pool_size=(2, 2), strides=(2, 2), padding='same'),keras.layers.Conv2D(filters=64,kernel_size=(5,5),padding="same",activation="relu"),keras.layers.MaxPooling2D(pool_size=(2, 2), strides=(2, 2), padding='same'),keras.layers.Conv2D(filters=128,kernel_size=(5,5),padding="same",activation="relu"),keras.layers.MaxPooling2D(pool_size=(2, 2), strides=(2, 2), padding='same'),keras.layers.Dropout(0.5),keras.layers.Flatten(),keras.layers.Dense(512,activation="relu"),keras.layers.Dense(self.dataset.num_classes,activation="softmax")])

模型訓(xùn)練+測試
其中epochs和batch_size可以自己選擇。

# 進(jìn)行模型訓(xùn)練的函數(shù)，具體的optimizer、loss可以進(jìn)行不同選擇def train_model(self):self.model.compile(optimizer = 'adam',loss = 'categorical_crossentropy',metrics=['accuracy'])# epochs、batch_size為可調(diào)的參數(shù)，epochs為訓(xùn)練多少輪、batch_size為每次訓(xùn)練多少個(gè)樣本self.model.fit(self.dataset.X_train,self.dataset.Y_train,epochs=50,batch_size=90)#用訓(xùn)練好的模型對測試集進(jìn)行測試def evaluate_model(self):print('Testing---------------')loss, accuracy = self.model.evaluate(self.dataset.X_test, self.dataset.Y_test)print('test loss:',loss)print('test accuracy:',accuracy)

訓(xùn)練結(jié)果如下所示:
(這準(zhǔn)確率高的我自己都害怕

Epoch 50/5090/2705 [..............................] - ETA: 1s - loss: 9.0821e-04 - accuracy: 1.0000270/2705 [=>............................] - ETA: 1s - loss: 5.1787e-04 - accuracy: 1.0000450/2705 [===>..........................] - ETA: 0s - loss: 0.0022 - accuracy: 0.9978 630/2705 [=====>........................] - ETA: 0s - loss: 0.0022 - accuracy: 0.9984810/2705 [=======>......................] - ETA: 0s - loss: 0.0023 - accuracy: 0.9988990/2705 [=========>....................] - ETA: 0s - loss: 0.0057 - accuracy: 0.9980 1170/2705 [===========>..................] - ETA: 0s - loss: 0.0055 - accuracy: 0.9983 1350/2705 [=============>................] - ETA: 0s - loss: 0.0049 - accuracy: 0.9985 1530/2705 [===============>..............] - ETA: 0s - loss: 0.0044 - accuracy: 0.9987 1710/2705 [=================>............] - ETA: 0s - loss: 0.0039 - accuracy: 0.9988 1890/2705 [===================>..........] - ETA: 0s - loss: 0.0036 - accuracy: 0.9989 2070/2705 [=====================>........] - ETA: 0s - loss: 0.0033 - accuracy: 0.9990 2250/2705 [=======================>......] - ETA: 0s - loss: 0.0031 - accuracy: 0.9991 2430/2705 [=========================>....] - ETA: 0s - loss: 0.0029 - accuracy: 0.9992 2610/2705 [===========================>..] - ETA: 0s - loss: 0.0027 - accuracy: 0.9992 2705/2705 [==============================] - 1s 443us/step - loss: 0.0026 - accuracy: 0.9993

測試集結(jié)果：
(攤手）

Testing---------------32/677 [>.............................] - ETA: 4s 192/677 [=======>......................] - ETA: 0s 352/677 [==============>...............] - ETA: 0s 512/677 [=====================>........] - ETA: 0s 672/677 [============================>.] - ETA: 0s 677/677 [==============================] - 0s 668us/step test loss: 0.00010534183920105803 test accuracy: 1.0

5.模型測試

對經(jīng)過訓(xùn)練之后的模型進(jìn)行測試，重新拍攝幾張照片進(jìn)行測試。

def test_onBatch(path):model= Model()model.load()index = 0img_list, label_lsit, counter = read_file(path)for img in img_list:picType,prob = model.predict(img)if picType != -1:index += 1name_list = read_name_list('E:\\faceRecognition')print(name_list[picType],prob)else:print(" Don't know this person")return index

結(jié)果如下所示：

Model Loaded. yqx_tst 1.0 yqx_tst 1.0 yqx_tst 1.0 yqx_tst 1.0 yqx_tst 1.0 yqx_tst 1.0 yqx_tst 1.0 yqx_tst 1.0 yqx_tst 1.0 yqx_tst 1.0

6.調(diào)用攝像頭識(shí)別

部分代碼：

def build_camera(self):# opencv文件中人臉級聯(lián)文件的位置，用于幫助識(shí)別圖像或者視頻流中的人臉face_cascade = cv2.CascadeClassifier('E:\ProgramData\Anaconda3\Lib\site-packages\cv2\data\haarcascade_frontalface_alt.xml')#讀取dataset數(shù)據(jù)集下的子文件夾名稱name_list = read_name_list("E:\\faceRecognition")#打開攝像頭并開始讀取畫面cameraCapture = cv2.VideoCapture(0,cv2.CAP_DSHOW)success,frame = cameraCapture.read()while success and cv2.waitKey(1) == -1:success,frame = cameraCapture.read()gray = cv2.cvtColor(frame,cv2.COLOR_BGR2GRAY)#圖像灰化faces = face_cascade.detectMultiScale(gray,1.3,5)for (x,y,w,h) in faces:ROI = gray[x:x+w,y:y+h]ROI = cv2.resize(ROI,(self.img_size,self.img_size),interpolation=cv2.INTER_LINEAR)label,prob = self.model.predict(ROI)#利用模型對cv2識(shí)別出來的人臉進(jìn)行對比if prob > 0.9:show_name = name_list[label]else:show_name = 'Stranger'cv2.putText(frame,show_name,(x,y-20),cv2.FONT_HERSHEY_SIMPLEX,1,255,2)#顯示名字frame = cv2.rectangle(frame,(x,y),(x+w,y+h),(255,0,0),2)#在人臉區(qū)域畫一個(gè)正方形出來cv2.imshow("Camera",frame)c = cv2.waitKey(10)if c & 0xFF == ord('q'):break# 釋放攝像頭并銷毀所有窗口cameraCapture.release()cv2.destroyAllWindows()

效果如下：



完整代碼請私信我。有不對的地方希望路過的大佬批評指正。

總結(jié)

以上是生活随笔為你收集整理的深度学习之基于opencv和CNN实现人脸识别的全部內(nèi)容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網(wǎng)站內(nèi)容還不錯(cuò)，歡迎將生活随笔推薦給好友。