视频帧输入 → SSD人脸检测 → 人脸ROI裁剪 → MobileNetV2口罩分类 → 结果标注输出

# 加载预训练MobileNetV2（不含顶层）baseModel = MobileNetV2(weights="imagenet", include_top=False,    input_tensor=Input(shape=(224, 224, 3)))# 构建自定义分类头headModel = baseModel.outputheadModel = AveragePooling2D(pool_size=(7, 7))(headModel)headModel = Flatten(name="flatten")(headModel)headModel = Dense(128, activation="relu")(headModel)headModel = Dropout(0.5)(headModel)headModel = Dense(2, activation="softmax")(headModel)# 组合完整模型model = Model(inputs=baseModel.input, outputs=headModel)# 冻结基础网络层for layer in baseModel.layers:    layer.trainable = False

# 加载SSD人脸检测器prototxtPath = "face_detector/deploy.prototxt"weightsPath = "face_detector/res10_300x300_ssd_iter_140000.caffemodel"faceNet = cv2.dnn.readNet(prototxtPath, weightsPath)# 检测人脸blob = cv2.dnn.blobFromImage(frame, 1.0, (300, 300), (104.0, 177.0, 123.0))faceNet.setInput(blob)detections = faceNet.forward()

# requirements.txttensorflow>=1.15.2keras==2.3.1imutils==0.5.3numpy==1.18.2opencv-python==4.2.0.*matplotlib==3.2.1scipy==1.4.1

pip install -r requirements.txt

dataset/├── with_mask/       # 戴口罩的人脸图片│   ├── 0_0_0 copy 10.jpg│   ├── 0_0_0 copy 11.jpg│   └── ...└── without_mask/    # 未戴口罩的人脸图片    ├── 0_0_caiguoqing_0130.jpg    ├── 0_0_caiyilin_0024.jpg    └── ...

aug = ImageDataGenerator(    rotation_range=20,        # 随机旋转±20度    zoom_range=0.15,          # 随机缩放±15%    width_shift_range=0.2,    # 水平平移±20%    height_shift_range=0.2,   # 垂直平移±20%    shear_range=0.15,         # 剪切变换±15%    horizontal_flip=True,     # 随机水平翻转    fill_mode="nearest"       # 填充模式)

# train_mask_detector.py 核心逻辑# 1. 加载数据data = []labels = []for category in ["with_mask", "without_mask"]:    path = os.path.join(DIRECTORY, category)    for img in os.listdir(path):        img_path = os.path.join(path, img)        image = load_img(img_path, target_size=(224, 224))        image = img_to_array(image)        image = preprocess_input(image)        data.append(image)        labels.append(category)# 2. 标签编码lb = LabelBinarizer()labels = lb.fit_transform(labels)labels = to_categorical(labels)# 3. 划分训练/测试集(trainX, testX, trainY, testY) = train_test_split(    data, labels, test_size=0.20, stratify=labels, random_state=42)# 4. 编译模型opt = Adam(lr=INIT_LR, decay=INIT_LR / EPOCHS)model.compile(loss="binary_crossentropy", optimizer=opt, metrics=["accuracy"])# 5. 训练H = model.fit(    aug.flow(trainX, trainY, batch_size=BS),    steps_per_epoch=len(trainX) // BS,    validation_data=(testX, testY),    validation_steps=len(testX) // BS,    epochs=EPOCHS)# 6. 保存模型model.save("mask_detector.model", save_format="h5")

# 在测试集上评估predIdxs = model.predict(testX, batch_size=BS)predIdxs = np.argmax(predIdxs, axis=1)# 输出分类报告print(classification_report(testY.argmax(axis=1), predIdxs,    target_names=lb.classes_))

def detect_and_predict_mask(frame, faceNet, maskNet):    # 1. 获取帧尺寸    (h, w) = frame.shape[:2]    # 2. 构建blob并进行人脸检测    blob = cv2.dnn.blobFromImage(frame, 1.0, (300, 300),         (104.0, 177.0, 123.0))    faceNet.setInput(blob)    detections = faceNet.forward()    faces = []    locs = []    preds = []    # 3. 遍历检测结果    for i in range(0, detections.shape[2]):        confidence = detections[0, 0, i, 2]        if confidence > 0.5:  # 置信度阈值            # 计算边界框坐标            box = detections[0, 0, i, 3:7] * np.array([w, h, w, h])            (startX, startY, endX, endY) = box.astype("int")            # 确保边界框在帧内            (startX, startY) = (max(0, startX), max(0, startY))            (endX, endY) = (min(w - 1, endX), min(h - 1, endY))            # 提取人脸ROI并预处理            face = frame[startY:endY, startX:endX]            face = cv2.cvtColor(face, cv2.COLOR_BGR2RGB)            face = cv2.resize(face, (224, 224))            face = img_to_array(face)            face = preprocess_input(face)            faces.append(face)            locs.append((startX, startY, endX, endY))    # 4. 批量预测    if len(faces) > 0:        faces = np.array(faces, dtype="float32")        preds = maskNet.predict(faces, batch_size=32)    return (locs, preds)

for (box, pred) in zip(locs, preds):    (startX, startY, endX, endY) = box    (mask, withoutMask) = pred    # 判断是否戴口罩    label = "Mask" if mask > withoutMask else "No Mask"    color = (0, 255, 0) if label == "Mask" else (0, 0, 255)    # 显示概率    label = "{}: {:.2f}%".format(label, max(mask, withoutMask) * 100)    # 绘制边界框和标签    cv2.putText(frame, label, (startX, startY - 10),        cv2.FONT_HERSHEY_SIMPLEX, 0.45, color, 2)    cv2.rectangle(frame, (startX, startY), (endX, endY), color, 2)

# 初始化视频流vs = VideoStream(src=0).start()while True:    frame = vs.read()    frame = imutils.resize(frame, width=400)    # 检测口罩    (locs, preds) = detect_and_predict_mask(frame, faceNet, maskNet)    # 可视化结果    for (box, pred) in zip(locs, preds):        # ... 绘制标注 ...    cv2.imshow("Frame", frame)    # 按'q'退出    if cv2.waitKey(1) & 0xFF == ord("q"):        breakcv2.destroyAllWindows()vs.stop()

Face-Mask-Detection-master/├── dataset/                          # 数据集│   ├── with_mask/                    # 戴口罩图片│   └── without_mask/                 # 未戴口罩图片├── face_detector/                    # 人脸检测模型│   ├── deploy.prototxt               # SSD网络配置│   └── res10_300x300_ssd_iter_140000.caffemodel  # 预训练权重├── train_mask_detector.py            # 训练脚本├── detect_mask_video.py              # 实时检测脚本├── mask_detector.model               # 训练好的口罩检测模型├── plot.png                          # 训练曲线图└── requirements.txt                  # 依赖列表

AttributeError: module 'tensorflow' has no attribute 'keras'

# 错误写法（TF 1.x风格）import keras# 正确写法（TF 2.x）from tensorflow import keras# 或from tensorflow.keras.models import load_model

error: (-215:Assertion failed) !_src.empty() in function 'cv::cvtColor'

# 尝试不同的摄像头索引vs = VideoStream(src=0).start()  # 尝试 0, 1, 2...# 或使用视频文件测试vs = cv2.VideoCapture("test_video.mp4")

OSError: Unable to open file (unable to open file: name = 'face_detector/deploy.prototxt')

import os# 使用绝对路径base_dir = os.path.dirname(os.path.abspath(__file__))prototxtPath = os.path.join(base_dir, "face_detector", "deploy.prototxt")

ResourceExhaustedError: OOM when allocating tensor

# 减小批次大小BS = 16  # 从32降到16# 或使用CPU训练import osos.environ["CUDA_VISIBLE_DEVICES"] = "-1"

# 解冻部分层进行微调（提升精度）for layer in baseModel.layers[-20:]:  # 解冻最后20层    layer.trainable = True# 使用更小的学习率进行微调opt = Adam(lr=INIT_LR / 10, decay=(INIT_LR / 10) / EPOCHS)