如何将图像分割为规则块并打乱重组？patchify的unpatchify报错排查

核心报错原因 #

np.random.shuffle()是原地打乱数组，执行后返回None，因此shuffled_patches实际为None，调用unpatchify时自然触发'NoneType' object has no attribute 'shape'错误。

此外还有两个潜在问题：

• 原图像尺寸需是patch大小的整数倍，否则patchify会自动截断边缘像素，后续unpatchify无法还原到原图像尺寸
• patchify返回的patch数组维度需要调整后才能正确打乱

修正后的完整代码 #

import numpy as np
from PIL import Image
from patchify import patchify, unpatchify
import matplotlib.pyplot as plt
import requests
import torchvision.transforms as transforms

# 预处理：统一图像尺寸为可被patch大小整除的数值（示例用200x200，对应100x100的patch）
transform = transforms.Compose(
    [transforms.Resize((200, 200)),
     transforms.ToTensor()])

url = 'https://hips.hearstapps.com/hmg-prod/images/bright-forget-me-nots-royalty-free-image-1677788394.jpg'
image = Image.open(requests.get(url, stream=True).raw)

# 转换为numpy数组（HWC格式）
img_np = np.array(transform(image).permute(1,2,0))

# 分割图像为100x100的patch
patch_size = (100, 100, 3)
patched_image = patchify(img_np, patch_size, step=100)
# patched_image形状为(2, 2, 1, 100, 100, 3)

# 调整维度，将patch展平为一维列表方便打乱
num_patches_h, num_patches_w = patched_image.shape[0], patched_image.shape[1]
patches_flat = patched_image.reshape(-1, *patch_size)  # 形状变为(4, 100, 100, 3)

# 生成打乱索引，避免原地修改导致的None问题
shuffled_indices = np.random.permutation(len(patches_flat))
shuffled_patches_flat = patches_flat[shuffled_indices]

# 恢复patch数组的原始维度结构
shuffled_patched = shuffled_patches_flat.reshape(num_patches_h, num_patches_w, 1, *patch_size)

# 拼接回完整图像
output_image = unpatchify(shuffled_patched, img_np.shape)

# 可视化对比
plt.figure(figsize=(12,6))
plt.subplot(121)
plt.title('Original Image')
plt.imshow(img_np)
plt.subplot(122)
plt.title('Shuffled Patches Image')
plt.imshow(output_image)
plt.show()

关键修正点说明 #

1. 解决NoneType问题：使用np.random.permutation()生成打乱索引，替代原地修改的shuffle()，确保得到有效数组
2. 适配patch尺寸：将图像resize为200x200（100的整数倍），避免patchify截断边缘，保证unpatchify能准确还原到原尺寸
3. 维度调整：先将patchify返回的多维度数组展平为一维patch列表，打乱后再恢复原patch数组的维度结构，满足unpatchify的输入格式要求

内容的提问来源于stack exchange，提问作者Rainbow