深度学习实践-使用CelebA_Spoof训练的权重测试NUAA

1.直接使用保存的网络测试NUAA

测试代码:

def read_test_file():
    base_path = r'E:\ml\fas\data\NUAA'
    val_file_path = os.path.join(base_path, "test.txt")
    train_image_path_list = []
    train_labels_list = []
    with open(val_file_path) as f:
        lines = f.readlines()
        for line in lines:
            image_path = line.split(',')[0]
            label = line.split(',')[1]
            img = cv2.imread(image_path)
            resize_img = cv2.resize(img, (100, 100))
            train_image_path_list.append(resize_img)
            train_labels_list.append(int(label))
    return np.asarray(train_image_path_list), np.asarray(train_labels_list)

def fit2():
    X_test, y_test = read_test_file()
    model = load_model('model/live_model.h5')
    test_loss, test_acc = model.evaluate(X_test, y_test, verbose=2)
    print(test_loss, test_acc)

if __name__ == '__main__':
    fit2()

测试结果

200/200 - 8s - loss: 1.3156 - accuracy: 0.5285
1.3155993223190308 0.5284998416900635

可以看到CelebA_Spoof 在自己的数据集测试可以达到99%以上
但是在NUAA数据集上测试准确率只能达到: 0.528
查看CelebA_Spoof 数据集后发现,CelebA_Spoof 的spoof数据集都是图像打印在纸上的加头像,但是NUAA里面是电脑屏幕或者手机屏幕录取的图片

2. 使用CelebA_Spoof 训练的网络再次训练NUAA数据

def read_train_file():
    base_path = r'E:\ml\fas\data\NUAA'
    train_file_path = os.path.join(base_path, "train.txt")

    train_image_path_list = []
    train_labels_list = []
    with open(train_file_path) as f:
        lines = f.readlines()
        for line in lines:
            image_path = line.split(',')[0]
            label = line.split(',')[1]
            img = cv2.imread(image_path)
            resize_img = cv2.resize(img, (100, 100))
            train_image_path_list.append(resize_img)
            train_labels_list.append(int(label))
    return np.asarray(train_image_path_list), np.asarray(train_labels_list)


def read_val_file():
    base_path = r'E:\ml\fas\data\NUAA'
    val_file_path = os.path.join(base_path, "val.txt")
    train_image_path_list = []
    train_labels_list = []
    with open(val_file_path) as f:
        lines = f.readlines()
        for line in lines:
            image_path = line.split(',')[0]
            label = line.split(',')[1]
            img = cv2.imread(image_path)
            resize_img = cv2.resize(img, (100, 100))
            train_image_path_list.append(resize_img)
            train_labels_list.append(int(label))
    return np.asarray(train_image_path_list), np.asarray(train_labels_list)


def read_test_file():
    base_path = r'E:\ml\fas\data\NUAA'
    val_file_path = os.path.join(base_path, "test.txt")
    train_image_path_list = []
    train_labels_list = []
    with open(val_file_path) as f:
        lines = f.readlines()
        for line in lines:
            image_path = line.split(',')[0]
            label = line.split(',')[1]
            img = cv2.imread(image_path)
            resize_img = cv2.resize(img, (100, 100))
            train_image_path_list.append(resize_img)
            train_labels_list.append(int(label))
    return np.asarray(train_image_path_list), np.asarray(train_labels_list)

def fit2():
    X_train, X_label = read_train_file()
    X_valid, y_valid = read_val_file()
    X_test, y_test = read_test_file()
    model = load_model('model/live_model.h5')
    history = model.fit(X_train, X_label, epochs=10, validation_data=(X_valid, y_valid))
    print(history)
    test_loss, test_acc = model.evaluate(X_test, y_test, verbose=2)
    print(test_loss, test_acc)


if __name__ == '__main__':
    fit2()

3.训练后测试结果

200/200 - 8s - loss: 5.3590 - accuracy: 0.6008
5.35904598236084 0.6008455753326416

虽然有些许提升,但是提升效果不怎么样