已开启
new: 新建文件 1班18.py #103
totm创建于 2025年12月29日
new: 新建文件 1班18.py #103
已开启
共 1 个文件变更+421-0
| @@ -0,0 +1,421 @@ | |||
| 1 | +import os | ||
| 2 | +import numpy as np | ||
| 3 | +import matplotlib.pyplot as plt | ||
| 4 | +import tensorflow as tf | ||
| 5 | +from tensorflow.keras import layers, models, optimizers, regularizers | ||
| 6 | +from tensorflow.keras.preprocessing.image import ImageDataGenerator | ||
| 7 | +from tensorflow.keras.callbacks import ModelCheckpoint, EarlyStopping, ReduceLROnPlateau | ||
| 8 | +from sklearn.model_selection import train_test_split | ||
| 9 | +from sklearn.preprocessing import LabelEncoder | ||
| 10 | +from sklearn.neighbors import KNeighborsClassifier | ||
| 11 | +from sklearn.svm import SVC | ||
| 12 | +from sklearn.metrics import classification_report, confusion_matrix | ||
| 13 | +import seaborn as sns | ||
| 14 | +import cv2 | ||
| 15 | +import mtcnn | ||
| 16 | +from tqdm import tqdm | ||
| 17 | + | ||
| 18 | +# 设置随机种子保证可重复性 | ||
| 19 | +np.random.seed(42) | ||
| 20 | +tf.random.set_seed(42) | ||
| 21 | + | ||
| 22 | +# 1. 数据准备和预处理 | ||
| 23 | +def load_data(data_dir, target_size=(160, 160)): | ||
| 24 | + """ | ||
| 25 | + 加载人脸数据集并进行预处理 | ||
| 26 | + """ | ||
| 27 | + images = [] | ||
| 28 | + labels = [] | ||
| 29 | + label_dict = {} | ||
| 30 | + current_label = 0 | ||
| 31 | + | ||
| 32 | + # 使用MTCNN进行人脸检测和对齐 | ||
| 33 | + detector = mtcnn.MTCNN() | ||
| 34 | + | ||
| 35 | + for person_name in tqdm(os.listdir(data_dir), desc="Loading data"): | ||
| 36 | + person_dir = os.path.join(data_dir, person_name) | ||
| 37 | + if not os.path.isdir(person_dir): | ||
| 38 | + continue | ||
| 39 | + | ||
| 40 | + label_dict[current_label] = person_name | ||
| 41 | + | ||
| 42 | + for image_name in os.listdir(person_dir): | ||
| 43 | + image_path = os.path.join(person_dir, image_name) | ||
| 44 | + try: | ||
| 45 | + # 读取图像 | ||
| 46 | + img = cv2.imread(image_path) | ||
| 47 | + img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) | ||
| 48 | + | ||
| 49 | + # 使用MTCNN检测人脸 | ||
| 50 | + results = detector.detect_faces(img) | ||
| 51 | + if len(results) == 0: | ||
| 52 | + continue | ||
| 53 | + | ||
| 54 | + # 获取第一个检测到的人脸 | ||
| 55 | + x1, y1, width, height = results[0]['box'] | ||
| 56 | + x2, y2 = x1 + width, y1 + height | ||
| 57 | + | ||
| 58 | + # 提取人脸区域 | ||
| 59 | + face = img[y1:y2, x1:x2] | ||
| 60 | + | ||
| 61 | + # 调整大小和预处理 | ||
| 62 | + face = cv2.resize(face, target_size) | ||
| 63 | + face = face.astype('float32') / 255.0 | ||
| 64 | + face = np.expand_dims(face, axis=0) | ||
| 65 | + | ||
| 66 | + # 数据增强 | ||
| 67 | + datagen = ImageDataGenerator( | ||
| 68 | + rotation_range=10, | ||
| 69 | + width_shift_range=0.1, | ||
| 70 | + height_shift_range=0.1, | ||
| 71 | + shear_range=0.1, | ||
| 72 | + zoom_range=0.1, | ||
| 73 | + horizontal_flip=True, | ||
| 74 | + fill_mode='nearest' | ||
| 75 | + ) | ||
| 76 | + | ||
| 77 | + # 应用数据增强并收集样本 | ||
| 78 | + aug_iter = datagen.flow(face, batch_size=1) | ||
| 79 | + for _ in range(5): # 每个原始图像生成5个增强样本 | ||
| 80 | + aug_img = next(aug_iter)[0].numpy() | ||
| 81 | + images.append(aug_img) | ||
| 82 | + labels.append(current_label) | ||
| 83 | + | ||
| 84 | + except Exception as e: | ||
| 85 | + print(f"Error processing {image_path}: {e}") | ||
| 86 | + continue | ||
| 87 | + | ||
| 88 | + current_label += 1 | ||
| 89 | + | ||
| 90 | + images = np.array(images) | ||
| 91 | + labels = np.array(labels) | ||
| 92 | + | ||
| 93 | + return images, labels, label_dict | ||
| 94 | + | ||
| 95 | +# 2. 构建FaceNet风格的模型 | ||
| 96 | +def build_facenet_model(input_shape=(160, 160, 3), embedding_size=128): | ||
| 97 | + """ | ||
| 98 | + 构建一个简化版的FaceNet模型 | ||
| 99 | + """ | ||
| 100 | + def inception_block(x, filters): | ||
| 101 | + """Inception模块""" | ||
| 102 | + path1 = layers.Conv2D(filters[0], 1, activation='relu')(x) | ||
| 103 | + | ||
| 104 | + path2 = layers.Conv2D(filters[1], 1, activation='relu')(x) | ||
| 105 | + path2 = layers.Conv2D(filters[1], 3, padding='same', activation='relu')(path2) | ||
| 106 | + | ||
| 107 | + path3 = layers.Conv2D(filters[2], 1, activation='relu')(x) | ||
| 108 | + path3 = layers.Conv2D(filters[2], 5, padding='same', activation='relu')(path3) | ||
| 109 | + | ||
| 110 | + path4 = layers.MaxPooling2D(3, 1, padding='same')(x) | ||
| 111 | + path4 = layers.Conv2D(filters[3], 1, activation='relu')(path4) | ||
| 112 | + | ||
| 113 | + return layers.Concatenate()([path1, path2, path3, path4]) | ||
| 114 | + | ||
| 115 | + # 输入层 | ||
| 116 | + inputs = layers.Input(shape=input_shape) | ||
| 117 | + x = layers.Lambda(lambda x: tf.image.per_image_standardization(x))(inputs) | ||
| 118 | + | ||
| 119 | + # 初始卷积层 | ||
| 120 | + x = layers.Conv2D(64, 7, strides=2, padding='same', activation='relu')(x) | ||
| 121 | + x = layers.MaxPooling2D(3, strides=2, padding='same')(x) | ||
| 122 | + | ||
| 123 | + # Inception模块堆叠 | ||
| 124 | + x = inception_block(x, [32, 32, 32, 32]) | ||
| 125 | + x = inception_block(x, [32, 32, 32, 32]) | ||
| 126 | + x = layers.MaxPooling2D(3, strides=2, padding='same')(x) | ||
| 127 | + | ||
| 128 | + x = inception_block(x, [64, 64, 64, 64]) | ||
| 129 | + x = inception_block(x, [64, 64, 64, 64]) | ||
| 130 | + x = layers.MaxPooling2D(3, strides=2, padding='same')(x) | ||
| 131 | + | ||
| 132 | + x = inception_block(x, [128, 128, 128, 128]) | ||
| 133 | + x = inception_block(x, [128, 128, 128, 128]) | ||
| 134 | + x = inception_block(x, [128, 128, 128, 128]) | ||
| 135 | + x = inception_block(x, [128, 128, 128, 128]) | ||
| 136 | + x = layers.MaxPooling2D(3, strides=2, padding='same')(x) | ||
| 137 | + | ||
| 138 | + # 展平和全连接层 | ||
| 139 | + x = layers.Flatten()(x) | ||
| 140 | + x = layers.Dense(256, activation='relu')(x) | ||
| 141 | + x = layers.Dropout(0.5)(x) | ||
| 142 | + | ||
| 143 | + # L2正则化嵌入层 | ||
| 144 | + embeddings = layers.Dense(embedding_size, activation=None, | ||
| 145 | + kernel_regularizer=regularizers.l2(0.01))(x) | ||
| 146 | + embeddings = tf.math.l2_normalize(embeddings, axis=1) | ||
| 147 | + | ||
| 148 | + # 构建模型 | ||
| 149 | + model = models.Model(inputs, embeddings, name='FaceNet') | ||
| 150 | + | ||
| 151 | + return model | ||
| 152 | + | ||
| 153 | +# 3. 三元组损失函数 | ||
| 154 | +def triplet_loss(y_true, y_pred, alpha=0.2): | ||
| 155 | + """ | ||
| 156 | + 三元组损失函数实现 | ||
| 157 | + """ | ||
| 158 | + anchor, positive, negative = y_pred[:, :128], y_pred[:, 128:256], y_pred[:, 256:] | ||
| 159 | + | ||
| 160 | + # 计算距离 | ||
| 161 | + pos_dist = tf.reduce_sum(tf.square(anchor - positive), axis=-1) | ||
| 162 | + neg_dist = tf.reduce_sum(tf.square(anchor - negative), axis=-1) | ||
| 163 | + | ||
| 164 | + # 计算基本损失 | ||
| 165 | + basic_loss = pos_dist - neg_dist + alpha | ||
| 166 | + loss = tf.reduce_mean(tf.maximum(basic_loss, 0.0)) | ||
| 167 | + | ||
| 168 | + return loss | ||
| 169 | + | ||
| 170 | +# 4. 数据生成器 | ||
| 171 | +class TripletGenerator: | ||
| 172 | + """ | ||
| 173 | + 生成三元组(anchor, positive, negative)的数据生成器 | ||
| 174 | + """ | ||
| 175 | + def __init__(self, images, labels, batch_size=32): | ||
| 176 | + self.images = images | ||
| 177 | + self.labels = labels | ||
| 178 | + self.batch_size = batch_size | ||
| 179 | + self.label_to_indices = {label: np.where(labels == label)[0] for label in np.unique(labels)} | ||
| 180 | + | ||
| 181 | + def __len__(self): | ||
| 182 | + return len(self.images) // self.batch_size | ||
| 183 | + | ||
| 184 | + def __getitem__(self, index): | ||
| 185 | + batch_indices = np.random.choice(len(self.images), self.batch_size * 3, replace=False) | ||
| 186 | + anchors, positives, negatives = [], [], [] | ||
| 187 | + | ||
| 188 | + for i in range(self.batch_size): | ||
| 189 | + # 随机选择一个anchor | ||
| 190 | + anchor_idx = batch_indices[i] | ||
| 191 | + anchor_label = self.labels[anchor_idx] | ||
| 192 | + | ||
| 193 | + # 选择同类的positive样本 | ||
| 194 | + pos_idx = np.random.choice(self.label_to_indices[anchor_label]) | ||
| 195 | + while pos_idx == anchor_idx: | ||
| 196 | + pos_idx = np.random.choice(self.label_to_indices[anchor_label]) | ||
| 197 | + | ||
| 198 | + # 选择不同类的negative样本 | ||
| 199 | + neg_label = np.random.choice([l for l in self.label_to_indices if l != anchor_label]) | ||
| 200 | + neg_idx = np.random.choice(self.label_to_indices[neg_label]) | ||
| 201 | + | ||
| 202 | + anchors.append(self.images[anchor_idx]) | ||
| 203 | + positives.append(self.images[pos_idx]) | ||
| 204 | + negatives.append(self.images[neg_idx]) | ||
| 205 | + | ||
| 206 | + anchors = np.array(anchors) | ||
| 207 | + positives = np.array(positives) | ||
| 208 | + negatives = np.array(negatives) | ||
| 209 | + | ||
| 210 | + # 组合成三元组 | ||
| 211 | + triplets = np.concatenate([anchors, positives, negatives], axis=-1) | ||
| 212 | + | ||
| 213 | + # 虚拟标签（三元组损失不需要真实标签） | ||
| 214 | + dummy_labels = np.zeros((self.batch_size, 1)) | ||
| 215 | + | ||
| 216 | + return triplets, dummy_labels | ||
| 217 | + | ||
| 218 | +# 5. 训练模型 | ||
| 219 | +def train_model(data_dir, model_save_path='facenet_model.h5'): | ||
| 220 | + # 加载和预处理数据 | ||
| 221 | + print("Loading and preprocessing data...") | ||
| 222 | + images, labels, label_dict = load_data(data_dir) | ||
| 223 | + | ||
| 224 | + # 划分训练集和验证集 | ||
| 225 | + X_train, X_val, y_train, y_val = train_test_split( | ||
| 226 | + images, labels, test_size=0.2, random_state=42, stratify=labels | ||
| 227 | + ) | ||
| 228 | + | ||
| 229 | + # 构建模型 | ||
| 230 | + print("Building model...") | ||
| 231 | + model = build_facenet_model() | ||
| 232 | + model.summary() | ||
| 233 | + | ||
| 234 | + # 编译模型（使用三元组损失） | ||
| 235 | + model.compile(optimizer=optimizers.Adam(learning_rate=0.001), loss=triplet_loss) | ||
| 236 | + | ||
| 237 | + # 回调函数 | ||
| 238 | + callbacks = [ | ||
| 239 | + ModelCheckpoint(model_save_path, monitor='val_loss', save_best_only=True), | ||
| 240 | + EarlyStopping(monitor='val_loss', patience=10, restore_best_weights=True), | ||
| 241 | + ReduceLROnPlateau(monitor='val_loss', factor=0.5, patience=3, min_lr=1e-6) | ||
| 242 | + ] | ||
| 243 | + | ||
| 244 | + # 数据生成器 | ||
| 245 | + train_generator = TripletGenerator(X_train, y_train, batch_size=32) | ||
| 246 | + val_generator = TripletGenerator(X_val, y_val, batch_size=32) | ||
| 247 | + | ||
| 248 | + # 训练模型 | ||
| 249 | + print("Training model...") | ||
| 250 | + history = model.fit( | ||
| 251 | + train_generator, | ||
| 252 | + validation_data=val_generator, | ||
| 253 | + epochs=50, | ||
| 254 | + callbacks=callbacks, | ||
| 255 | + verbose=1 | ||
| 256 | + ) | ||
| 257 | + | ||
| 258 | + # 绘制训练曲线 | ||
| 259 | + plt.figure(figsize=(12, 5)) | ||
| 260 | + plt.subplot(1, 2, 1) | ||
| 261 | + plt.plot(history.history['loss'], label='Train Loss') | ||
| 262 | + plt.plot(history.history['val_loss'], label='Validation Loss') | ||
| 263 | + plt.title('Model Loss') | ||
| 264 | + plt.xlabel('Epoch') | ||
| 265 | + plt.ylabel('Loss') | ||
| 266 | + plt.legend() | ||
| 267 | + | ||
| 268 | + plt.subplot(1, 2, 2) | ||
| 269 | + plt.plot(history.history['lr'], label='Learning Rate') | ||
| 270 | + plt.title('Learning Rate Schedule') | ||
| 271 | + plt.xlabel('Epoch') | ||
| 272 | + plt.ylabel('Learning Rate') | ||
| 273 | + plt.legend() | ||
| 274 | + | ||
| 275 | + plt.tight_layout() | ||
| 276 | + plt.savefig('training_curves.png') | ||
| 277 | + plt.show() | ||
| 278 | + | ||
| 279 | + return model, label_dict | ||
| 280 | + | ||
| 281 | +# 6. 评估模型 | ||
| 282 | +def evaluate_model(model, images, labels, label_dict): | ||
| 283 | + """ | ||
| 284 | + 评估模型性能 | ||
| 285 | + """ | ||
| 286 | + # 提取嵌入向量 | ||
| 287 | + print("Extracting embeddings...") | ||
| 288 | + embeddings = model.predict(images, batch_size=32, verbose=1) | ||
| 289 | + | ||
| 290 | + # 编码标签 | ||
| 291 | + le = LabelEncoder() | ||
| 292 | + y_true = le.fit_transform(labels) | ||
| 293 | + num_classes = len(le.classes_) | ||
| 294 | + | ||
| 295 | + # 训练分类器（SVM和KNN） | ||
| 296 | + print("Training classifiers...") | ||
| 297 | + X_train, X_test, y_train, y_test = train_test_split( | ||
| 298 | + embeddings, y_true, test_size=0.2, random_state=42, stratify=y_true | ||
| 299 | + ) | ||
| 300 | + | ||
| 301 | + # SVM分类器 | ||
| 302 | + svm = SVC(kernel='rbf', probability=True) | ||
| 303 | + svm.fit(X_train, y_train) | ||
| 304 | + y_pred_svm = svm.predict(X_test) | ||
| 305 | + | ||
| 306 | + # KNN分类器 | ||
| 307 | + knn = KNeighborsClassifier(n_neighbors=5) | ||
| 308 | + knn.fit(X_train, y_train) | ||
| 309 | + y_pred_knn = knn.predict(X_test) | ||
| 310 | + | ||
| 311 | + # 评估结果 | ||
| 312 | + print("\nSVM Classification Report:") | ||
| 313 | + print(classification_report(y_test, y_pred_svm, target_names=le.classes_)) | ||
| 314 | + | ||
| 315 | + print("\nKNN Classification Report:") | ||
| 316 | + print(classification_report(y_test, y_pred_knn, target_names=le.classes_)) | ||
| 317 | + | ||
| 318 | + # 混淆矩阵 | ||
| 319 | + plt.figure(figsize=(12, 6)) | ||
| 320 | + plt.subplot(1, 2, 1) | ||
| 321 | + cm = confusion_matrix(y_test, y_pred_svm) | ||
| 322 | + sns.heatmap(cm, annot=True, fmt='d', cmap='Blues', | ||
| 323 | + xticklabels=le.classes_, yticklabels=le.classes_) | ||
| 324 | + plt.title('SVM Confusion Matrix') | ||
| 325 | + | ||
| 326 | + plt.subplot(1, 2, 2) | ||
| 327 | + cm = confusion_matrix(y_test, y_pred_knn) | ||
| 328 | + sns.heatmap(cm, annot=True, fmt='d', cmap='Greens', | ||
| 329 | + xticklabels=le.classes_, yticklabels=le.classes_) | ||
| 330 | + plt.title('KNN Confusion Matrix') | ||
| 331 | + | ||
| 332 | + plt.tight_layout() | ||
| 333 | + plt.savefig('confusion_matrices.png') | ||
| 334 | + plt.show() | ||
| 335 | + | ||
| 336 | + return svm, knn, le | ||
| 337 | + | ||
| 338 | +# 7. 实时人脸识别 | ||
| 339 | +def real_time_recognition(model, classifier, le, label_dict, threshold=0.7): | ||
| 340 | + """ | ||
| 341 | + 实时人脸识别演示 | ||
| 342 | + """ | ||
| 343 | + # 加载检测器 | ||
| 344 | + detector = mtcnn.MTCNN() | ||
| 345 | + | ||
| 346 | + # 打开摄像头 | ||
| 347 | + cap = cv2.VideoCapture(0) | ||
| 348 | + | ||
| 349 | + while True: | ||
| 350 | + ret, frame = cap.read() | ||
| 351 | + if not ret: | ||
| 352 | + break | ||
| 353 | + | ||
| 354 | + # 检测人脸 | ||
| 355 | + rgb_frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB) | ||
| 356 | + results = detector.detect_faces(rgb_frame) | ||
| 357 | + | ||
| 358 | + for result in results: | ||
| 359 | + x1, y1, width, height = result['box'] | ||
| 360 | + x2, y2 = x1 + width, y1 + height | ||
| 361 | + | ||
| 362 | + # 提取人脸并预处理 | ||
| 363 | + face = rgb_frame[y1:y2, x1:x2] | ||
| 364 | + face = cv2.resize(face, (160, 160)) | ||
| 365 | + face = face.astype('float32') / 255.0 | ||
| 366 | + face = np.expand_dims(face, axis=0) | ||
| 367 | + | ||
| 368 | + # 预测嵌入向量 | ||
| 369 | + embedding = model.predict(face)[0] | ||
| 370 | + | ||
| 371 | + # 预测类别 | ||
| 372 | + if isinstance(classifier, SVC): | ||
| 373 | + probas = classifier.predict_proba([embedding])[0] | ||
| 374 | + max_prob = np.max(probas) | ||
| 375 | + pred_idx = np.argmax(probas) | ||
| 376 | + else: # KNN | ||
| 377 | + distances, indices = classifier.kneighbors([embedding]) | ||
| 378 | + max_prob = 1 / (distances[0][0] + 1e-6) # 转换为类似概率的值 | ||
| 379 | + pred_idx = indices[0][0] | ||
| 380 | + | ||
| 381 | + # 获取预测标签 | ||
| 382 | + pred_label = le.inverse_transform([pred_idx])[0] | ||
| 383 | + person_name = label_dict.get(pred_label, "Unknown") | ||
| 384 | + | ||
| 385 | + # 绘制边界框和标签 | ||
| 386 | + color = (0, 255, 0) if max_prob >= threshold else (0, 0, 255) | ||
| 387 | + label = f"{person_name} ({max_prob:.2f})" | ||
| 388 | + | ||
| 389 | + cv2.rectangle(frame, (x1, y1), (x2, y2), color, 2) | ||
| 390 | + cv2.putText(frame, label, (x1 y1 - 10), | ||
| 391 | + cv2.FONT_HERSHEY_SIMPLEX, 0.5, color, 2) | ||
| 392 | + | ||
| 393 | + cv2.imshow('Real-time Face Recognition', frame) | ||
| 394 | + | ||
| 395 | + if cv2.waitKey(1) & 0xFF == ord('q'): | ||
| 396 | + break | ||
| 397 | + | ||
| 398 | + cap.release() | ||
| 399 | + cv2.destroyAllWindows() | ||
| 400 | + | ||
| 401 | +# 主函数 | ||
| 402 | +def main(): | ||
| 403 | + # 数据目录（应包含按人物命名的子目录，每个子目录包含该人物的多张照片） | ||
| 404 | + data_dir = 'lfw_dataset' # 替换为你的数据集路径 | ||
| 405 | + | ||
| 406 | + # 1. 训练模型 | ||
| 407 | + print("Starting Face Recognition System Training...") | ||
| 408 | + model, label_dict = train_model(data_dir) | ||
| 409 | + | ||
| 410 | + # 2. 加载最佳模型并评估 | ||
| 411 | + print("\nEvaluating Model...") | ||
| 412 | + images, labels, _ = load_data(data_dir) # 重新加载数据 | ||
| 413 | + svm, knn, le = evaluate_model(model, images, labels, label_dict) | ||
| 414 | + | ||
| 415 | + # 3. 实时识别演示 | ||
| 416 | + print("\nStarting Real-time Recognition...") | ||
| 417 | + print("Press 'q' to quit.") | ||
| 418 | + real_time_recognition(model, svm, le, label_dict) | ||
| 419 | + | ||
| 420 | +if __name__ == '__main__': | ||
| 421 | + main() | ||