import tensorflow as tf
from tensorflow.keras.models import load_model, Sequential
from tensorflow.keras.layers import Conv2D, MaxPooling2D, Flatten, Dense, Dropout
from tensorflow.keras.preprocessing.image import ImageDataGenerator

def create_model():
    model = Sequential([
        Conv2D(32, (3, 3), activation='relu', input_shape=(640, 640, 3)),
        MaxPooling2D((2, 2)),
        Conv2D(64, (3, 3), activation='relu'),
        MaxPooling2D((2, 2)),
        Flatten(),
        Dense(128, activation='relu'),
        Dropout(0.5),
        Dense(10, activation='softmax')  # 假设有10个交通标志类别
    ])
    model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])
    return model

def train_model(model, train_data_dir, validation_data_dir):
    train_datagen = ImageDataGenerator(rescale=1./255)
    val_datagen = ImageDataGenerator(rescale=1./255)

    train_generator = train_datagen.flow_from_directory(
        train_data_dir,
        target_size=(640, 640),
        batch_size=32,
        class_mode='sparse'
    )

    validation_generator = val_datagen.flow_from_directory(
        validation_data_dir,
        target_size=(640, 640),
        batch_size=32,
        class_mode='sparse'
    )

    model.fit(train_generator, epochs=10, validation_data=validation_generator)

# 示例
model = create_model()
train_data_dir = 'path/to/train/data'
validation_data_dir = 'path/to/validation/data'
train_model(model, train_data_dir, validation_data_dir)

# 保存模型
model.save('path/to/your/model.h5')

# 加载模型并预测
model = load_model('path/to/your/model.h5')

def predict_traffic_sign(image):
    # 扩展维度以适配模型输入
    image = np.expand_dims(image, axis=0)
    # 模型预测
    predictions = model.predict(image)
    return predictions

# 示例
predictions = predict_traffic_sign(preprocessed_image)
predicted_class = np.argmax(predictions, axis=1)