基于CycleGAN-and-pix2pix的单图推理结果模糊问题求助

我在Google Colab上基于CycleGAN-and-pix2pix开源API训练了CycleGAN模型，训练命令为：

!python train.py --dataroot /content/drive/MyDrive/project/dataset --name F2F --model cycle_gan --display_id -1

通过数据加载器从文件夹批量推理的代码运行正常，生成结果良好，但自行编写单图推理的预处理函数后，生成的图像非常模糊，达不到批量推理的效果，希望解决这个问题。

批量推理代码 #

opt = TestOptions()
# defined options occurs here
dataset = create_dataset(opt)  
# Initialize the model
model = create_model(opt)
model.setup(opt)
model.eval()
data_iter = iter(dataset.dataloader)
data_dict = next(data_iter)
input_image_tensor = data_dict['A']
data = {'A': input_image_tensor,'A_paths': ''}
model.set_input(data)
model.test()
visuals = model.get_current_visuals()
output_image = visuals['fake']
output_image_np = output_image.squeeze().cpu().numpy().transpose(1, 2, 0)
output_image_np = ((output_image_np - output_image_np.min()) / (output_image_np.max() - output_image_np.min()) * 255).astype(np.uint8)
output_image_np = cv2.cvtColor(output_image_np, cv2.COLOR_BGR2RGB)
cv2_imshow(output_image_np)

自定义预处理及单图推理代码 #

预处理函数 #

def preprocess(image):
    if image.ndim == 2 or image.shape[2] == 1:
        image = cv2.cvtColor(image, cv2.COLOR_GRAY2RGB)
    elif image.shape[2] == 4:
        image = cv2.cvtColor(image, cv2.COLOR_BGRA2BGR)
    elif image.shape[2] == 3:
        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
       image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
    pil_image = transforms.ToPILImage()(image)
    transform_pipeline = transforms.Compose([
        transforms.Resize(286),
        transforms.CenterCrop(256),
        transforms.ToTensor(),
        transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
    ])
       image_tensor = transform_pipeline(pil_image)
       image_tensor = image_tensor.unsqueeze(0)

    return image_tensor

单图推理代码 #

input_image = cv2.imread('/content/drive/MyDrive/dataset/testA/image_1.jpg')
input_image_tensor = preprocess(input_image)
data = {'A': input_image_tensor,'A_paths': ''}
model.set_input(data)
model.test()
visuals = model.get_current_visuals()
output_image = visuals['fake']
output_image_np = output_image.squeeze().cpu().numpy().transpose(1, 2, 0)
output_image_np = ((output_image_np - output_image_np.min()) / (output_image_np.max() - output_image_np.min()) * 255).astype(np.uint8)
output_image_np = cv2.cvtColor(output_image_np, cv2.COLOR_BGR2RGB)
cv2_imshow(output_image_np)

解决方案 #

问题定位 #

原预处理函数存在两个关键问题：

1. 重复颜色空间转换：在处理3通道图像时，连续执行了两次cv2.COLOR_BGR2RGB转换，导致原本正确的RGB图像被再次反转，颜色空间错乱，模型接收到异常输入后生成模糊结果。
2. 插值方式不明确：transforms.Resize未显式指定插值方式，虽然默认是双线性，但为了和批量推理时数据集的预处理逻辑完全对齐，需要明确指定。

修正后的预处理函数 #

def preprocess(image):
    # 处理不同通道数的图像
    if image.ndim == 2 or image.shape[2] == 1:
        image = cv2.cvtColor(image, cv2.COLOR_GRAY2RGB)
    elif image.shape[2] == 4:
        image = cv2.cvtColor(image, cv2.COLOR_BGRA2BGR)
    elif image.shape[2] == 3:
        # 仅执行一次BGR到RGB的转换
        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
    
    pil_image = transforms.ToPILImage()(image)
    # 显式指定双线性插值，与CycleGAN默认预处理逻辑一致
    transform_pipeline = transforms.Compose([
        transforms.Resize(286, interpolation=transforms.InterpolationMode.BILINEAR),
        transforms.CenterCrop(256),
        transforms.ToTensor(),
        transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
    ])
    image_tensor = transform_pipeline(pil_image)
    image_tensor = image_tensor.unsqueeze(0)

    return image_tensor

额外验证建议 #

• 检查输入图像的宽高比，确保经过Resize(286)和CenterCrop(256)后没有出现严重变形，若原图比例特殊，可调整预处理逻辑适配。
• 对比批量推理时输入张量和单图预处理后的张量，查看数值范围、维度是否完全一致，进一步排查输入差异。

内容的提问来源于stack exchange，提问作者yair yatzkan