homework4

homework4/face.keypoints.py

+import numpy as np
+import time
+import os
+import cv2
+import kmodel
+from utils import transparentOverlay
+
+os.environ['KMP_DUPLICATE_LIB_OK']='True'
+
+# 加载预先训练好的模型
+# my_model = kmodel.load_trained_model('yuan_model_mac')
+# 加载自己训练好的模型（测试时取消下面行的注释）
+my_model = kmodel.load_trained_model('my_model')
+
+# 创建人脸检测器
+face_cascade = cv2.CascadeClassifier('cascades/haarcascade_frontalface_default.xml')
+
+# 加载摄像头
+camera = cv2.VideoCapture(0)
+
+# 加载一个太阳眼镜图像
+sunglasses = cv2.imread('sunglass.png', cv2.IMREAD_UNCHANGED)
+
+# 死循环
+while True:
+    time.sleep(0.1)
+
+    # 从摄像头获取一张图像
+    (_, frame) = camera.read()
+    frame = cv2.flip(frame, 1)
+    frame2 = np.copy(frame)
+    gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
+
+    # 检测所有的人脸
+    faces = face_cascade.detectMultiScale(gray, 1.25, 6)
+
+    # 对每一个检测到的人脸
+    for (x, y, w, h) in faces:
+
+        # 只包含人脸的图像
+        gray_face = gray[y:y+h, x:x+w]
+        color_face = frame[y:y+h, x:x+w]
+
+        # 将人脸图像的值 normalize 在 [0, 1] 之间
+        gray_normalized = gray_face / 255
+
+        # 缩放灰度图人脸到 96x96 匹配网络的输入
+        original_shape = gray_face.shape # A Copy for future reference
+        face_resized = cv2.resize(gray_normalized, (96, 96), interpolation = cv2.INTER_AREA)
+        face_resized = face_resized.reshape(1, 96, 96, 1)
+
+        # 预测关键点坐标
+        keypoints = my_model.predict(face_resized)
+
+        # 将关键点坐标的值从 [-1, 1] 之间转换为 [0, 96] 之间
+        keypoints = keypoints * 48 + 48
+
+        # 缩放彩色图人脸到 96x96 匹配关键点
+        face_resized_color = cv2.resize(color_face, (96, 96), interpolation = cv2.INTER_AREA)
+        face_resized_color2 = np.copy(face_resized_color)
+
+        # 将网络输出的30个值配对为15个tuple对
+        points = []
+        for i, co in enumerate(keypoints[0][0::2]):
+            points.append((co, keypoints[0][1::2][i]))
+
+        # 按照关键点的 left_eyebrow_outer_end_x[7], right_eyebrow_outer_end_x[9]确定眼镜的宽度
+        sunglass_width = int((points[7][0]-points[9][0])*1.1)
+
+        # 按照关键点的 nose_tip_y[10], right_eyebrow_inner_end_y[8]确定眼镜的高度
+        sunglass_height = int((points[10][1]-points[8][1])/1.1)
+        sunglass_resized = cv2.resize(sunglasses, (sunglass_width, sunglass_height), interpolation = cv2.INTER_CUBIC)
+        face_resized_color = transparentOverlay(face_resized_color, sunglass_resized , pos=(int(points[9][0]),int(points[9][1])), scale = 1)
+
+        # 将覆盖了眼镜的 face_resized_color 图像转为摄像头捕捉到的原始图像中的大小
+        frame[y:y+h, x:x+w] = cv2.resize(face_resized_color, original_shape, interpolation = cv2.INTER_CUBIC)
+
+        # 在人脸图像中显示关键点坐标
+        for keypoint in points:
+            cv2.circle(face_resized_color2, keypoint, 1, (0,255,0), 1)
+
+        frame2[y:y+h, x:x+w] = cv2.resize(face_resized_color2, original_shape, interpolation = cv2.INTER_CUBIC)
+
+        # 显示加了眼镜的图像
+        cv2.imshow("With Glass", frame)
+        # 显示添加了关键点的图像
+        cv2.imshow("With Keypoints", frame2)
+
+    # 当 'q' 键被点击, 退出循环
+    if cv2.waitKey(1) & 0xFF == ord("q"):
+        break
+
+# 释放摄像头, 关闭窗口
+camera.release()
+cv2.destroyAllWindows()

homework4/kmodel.py

+from keras.models import Sequential
+from keras.models import load_model
+from keras.layers import Activation, BatchNormalization, Convolution2D, \
+    MaxPooling2D, Dropout, Input
+from keras.layers import Flatten, Dense
+from keras.optimizers import SGD, RMSprop, Adagrad, Adadelta, Adam, Adamax, Nadam
+
+def create_model():
+    '''
+    网络的输入为96x96的单通道灰阶图像, 输出30个值, 代表的15个关键点的横坐标和纵坐标
+    '''
+    model = Sequential()
+    # 96
+    model.add(BatchNormalization(input_shape=(96, 96, 1)))
+    model.add(Convolution2D(32, (5, 5), activation='relu', padding="same"))
+    model.add(MaxPooling2D(pool_size=(2, 2)))
+
+    planes = 64
+    # 48 -> 24 -> 12 -> 6 -> 3
+    for _ in range(4):
+        model.add(Convolution2D(planes, (3, 3), padding="same"))
+        model.add(BatchNormalization())
+        model.add(Activation(activation="relu"))
+        model.add(Convolution2D(planes, (3, 3), padding="same"))
+        model.add(BatchNormalization())
+        model.add(MaxPooling2D(pool_size=(2, 2)))
+        planes *= 2
+
+    # 1
+    model.add(Convolution2D(planes, (3, 3), padding="valid", activation="relu"))
+    model.add(Convolution2D(planes // 2, (1, 1), padding="valid", activation="relu"))
+    model.add(Flatten())
+    model.add(Dense(planes // 4, activation="relu"))
+    model.add(Dropout(rate=0.5))
+    model.add(Dense(30, activation="tanh"))
+
+    return model
+
+def compile_model(model):
+    optimizer = Adam(lr=1e-3)
+    loss = "mean_squared_error"
+    metrics = ['mae']
+    model.compile(optimizer=optimizer, loss=loss, metrics=metrics)
+
+def train_model(model, X_train, y_train):
+    return model.fit(X_train, y_train, epochs=500, batch_size=200, verbose=1, validation_split=0.2)
+
+def save_model(model, fileName):
+    model.save(fileName + '.h5')
+
+def load_trained_model(fileName):
+    return load_model(fileName + '.h5')

homework4/model_builder.py

+from utils import load_data
+import kmodel
+import sys
+import pickle
+
+# 加载训练数据
+X_train, y_train = load_data()
+
+# 创建网络结构
+my_model = kmodel.create_model()
+with open("net.txt", mode="w") as f:
+    orig_stdout = sys.stdout
+    sys.stdout = f
+    my_model.summary()
+    sys.stdout = orig_stdout
+
+# 编译网络模型
+kmodel.compile_model(my_model)
+
+# my_model = kmodel.load_trained_model('my_model')
+# 训练网络模型
+history = kmodel.train_model(my_model, X_train, y_train)
+with open("log.pickle", mode="wb") as f:
+    pickle.dump(history.history, f)
+
+# 保存网络模型
+kmodel.save_model(my_model, 'my_model')
+
+
+

homework4/net.txt

+_________________________________________________________________
+Layer (type)                 Output Shape              Param #   
+=================================================================
+batch_normalization_1 (Batch (None, 96, 96, 1)         4         
+_________________________________________________________________
+conv2d_1 (Conv2D)            (None, 96, 96, 32)        832       
+_________________________________________________________________
+max_pooling2d_1 (MaxPooling2 (None, 48, 48, 32)        0         
+_________________________________________________________________
+conv2d_2 (Conv2D)            (None, 48, 48, 64)        18496     
+_________________________________________________________________
+batch_normalization_2 (Batch (None, 48, 48, 64)        256       
+_________________________________________________________________
+activation_1 (Activation)    (None, 48, 48, 64)        0         
+_________________________________________________________________
+conv2d_3 (Conv2D)            (None, 48, 48, 64)        36928     
+_________________________________________________________________
+batch_normalization_3 (Batch (None, 48, 48, 64)        256       
+_________________________________________________________________
+max_pooling2d_2 (MaxPooling2 (None, 24, 24, 64)        0         
+_________________________________________________________________
+conv2d_4 (Conv2D)            (None, 24, 24, 128)       73856     
+_________________________________________________________________
+batch_normalization_4 (Batch (None, 24, 24, 128)       512       
+_________________________________________________________________
+activation_2 (Activation)    (None, 24, 24, 128)       0         
+_________________________________________________________________
+conv2d_5 (Conv2D)            (None, 24, 24, 128)       147584    
+_________________________________________________________________
+batch_normalization_5 (Batch (None, 24, 24, 128)       512       
+_________________________________________________________________
+max_pooling2d_3 (MaxPooling2 (None, 12, 12, 128)       0         
+_________________________________________________________________
+conv2d_6 (Conv2D)            (None, 12, 12, 256)       295168    
+_________________________________________________________________
+batch_normalization_6 (Batch (None, 12, 12, 256)       1024      
+_________________________________________________________________
+activation_3 (Activation)    (None, 12, 12, 256)       0         
+_________________________________________________________________
+conv2d_7 (Conv2D)            (None, 12, 12, 256)       590080    
+_________________________________________________________________
+batch_normalization_7 (Batch (None, 12, 12, 256)       1024      
+_________________________________________________________________
+max_pooling2d_4 (MaxPooling2 (None, 6, 6, 256)         0         
+_________________________________________________________________
+conv2d_8 (Conv2D)            (None, 6, 6, 512)         1180160   
+_________________________________________________________________
+batch_normalization_8 (Batch (None, 6, 6, 512)         2048      
+_________________________________________________________________
+activation_4 (Activation)    (None, 6, 6, 512)         0         
+_________________________________________________________________
+conv2d_9 (Conv2D)            (None, 6, 6, 512)         2359808   
+_________________________________________________________________
+batch_normalization_9 (Batch (None, 6, 6, 512)         2048      
+_________________________________________________________________
+max_pooling2d_5 (MaxPooling2 (None, 3, 3, 512)         0         
+_________________________________________________________________
+conv2d_10 (Conv2D)           (None, 1, 1, 1024)        4719616   
+_________________________________________________________________
+conv2d_11 (Conv2D)           (None, 1, 1, 512)         524800    
+_________________________________________________________________
+flatten_1 (Flatten)          (None, 512)               0         
+_________________________________________________________________
+dense_1 (Dense)              (None, 256)               131328    
+_________________________________________________________________
+dropout_1 (Dropout)          (None, 256)               0         
+_________________________________________________________________
+dense_2 (Dense)              (None, 30)                7710      
+=================================================================
+Total params: 10,094,050
+Trainable params: 10,090,208
+Non-trainable params: 3,842
+_________________________________________________________________

homework4/utils.py

+import numpy as np
+from pandas.io.parsers import read_csv
+from sklearn.utils import shuffle
+import cv2
+
+def load_data(test=False):
+    """
+    当 test 为真, 加载测试数据, 否则加载训练数据 
+    """
+    FTRAIN = './data/training.csv'
+    FTEST = './data/test.csv'
+    fname = FTEST if test else FTRAIN
+    df = read_csv(fname)
+
+    # 将'Image' 列中 '空白键' 分割的数字们转换为一个 numpy array
+    df['Image'] = df['Image'].apply(lambda im: np.fromstring(im, sep=' '))
+
+    # 丢弃有缺失值的数据
+    df = df.dropna()  
+
+    # 将图像的数字从 0 到 255 的整数转换为 0 到 1 的实数
+    X = np.vstack(df['Image'].values) / 255.  
+    X = X.astype(np.float32)
+
+    # 将 X 的每一行转换为一个 96 * 96 * 1 的三维数组
+    X = X.reshape(-1, 96, 96, 1) 
+
+    # 只有 FTRAIN 包含关键点的数据 (target value)
+    if not test:  
+        y = df[df.columns[:-1]].values
+        # 将关键点的值 normalize 到 [-1, 1] 之间
+        y = (y - 48) / 48  
+        # 置乱训练数据
+        X, y = shuffle(X, y, random_state=42)  
+        y = y.astype(np.float32)
+    else:
+        y = None
+
+    return X, y
+
+def transparentOverlay(src , overlay , pos=(0,0), scale = 1):
+    """
+    将带透明通道(png图像)的图像 overlay 叠放在src图像上方
+    :param src: 背景图像
+    :param overlay: 带透明通道的图像 (BGRA)
+    :param pos: 叠放的起始位置
+    :param scale : overlay图像的缩放因子
+    :return: Resultant Image
+    """
+    if scale != 1:
+        overlay = cv2.resize(overlay,(0,0),fx=scale,fy=scale)
+
+    # overlay图像的高和宽
+    h,w,_ = overlay.shape  
+    # 叠放的起始坐标
+    y,x = pos[0],pos[1] 
+    
+    # 以下被注释的代码是没有优化的版本, 便于理解, 与如下没有注释的版本的功能一样
+    """     
+    # src图像的高和款
+    rows,cols,_ = src.shape  
+    for i in range(h):
+        for j in range(w):
+            if x+i >= rows or y+j >= cols:
+                continue
+            alpha = float(overlay[i][j][3]/255.0) # 读取alpha通道的值
+            src[x+i][y+j] = alpha*overlay[i][j][:3]+(1-alpha)*src[x+i][y+j]
+    return src """
+
+    alpha = overlay[:,:,3]/255.0
+    alpha = alpha[..., np.newaxis]
+    src[x:x+h,y:y+w,:] = alpha * overlay[:,:,:3] + (1-alpha)*src[x:x+h,y:y+w,:]
+    return src
\ No newline at end of file