将Base64编码的图像转换为numpy数组

Question

I'm receiving an image via an http POST that is base64 encoded. It can be a JPG or a BMP. Now that I have the image, I can get it in memory. I found how to write it to disk and re-read it into a numpy array (Which I need to actually put into a torch.tensor but numpy will suffice for now).

HEre's what works for me but appears HIGHLY inefficient :

import torch
import numpy as np
from PIL import Image
import base64

base64_decoded = base64.b64decode(test_image_base64_encoded)

with open("out.jpg", "wb") as out_file:
    out_file.write(base64_decoded)

image = Image.open("out.jpg")
image_np = np.array(image)
image_torch = torch.tensor(np.array(image))

It feels extremely useless to have to write the array to out.jpg to reread it right after into an array. There must be a better way. I've tried some things where it ends up in a 1D array... and my image is a 2D array in my case (BW image).

nparr = np.fromstring(base64.b64decode(test_image), np.uint8)

would yield when nparr.shape = (694463,) when image_np.shape = (2048, 3072)

Any idea how I could represent to np.array something like Image.frombase64 :) ? I know it doesn't exists per say but that would be great if it could somehow interpret the "file" without having to save it to disk first.

Answer 1

Assuming you're using PIL, but you don't know the image type or dimensions:

from PIL import Image
import base64
import io
import numpy as np
import torch

base64_decoded = base64.b64decode(test_image_base64_encoded)

image = Image.open(io.BytesIO(base64_decoded))
image_np = np.array(image)
image_torch = torch.tensor(np.array(image))

io.BytesIO is the key thing you're missing, I think.