I have a string Tensor object that I produced by calling tf.io.encode_jpeg on an image Tensor (Documentation for encode_jpeg).

How do I convert this string Tensor into a PIL Image?

I've tried calling Image.fromarray(encoded_tensor.numpy()), but this returns AttributeError: 'bytes' object has no attribute '__array_interface__'.

CodePudding user response：

You appear to have an in-memory (rather than on-disk) JPEG-encoded image. You can check if that is correct by printing the start of the buffer:

print(encoded_tensor[:10])

and seeing if it starts with the JPEG magic number ff d8 ff.

If so, you need to wrap it in a BytesIO and open it into a PIL Image with:

from io import BytesIO
from PIL import Image

im = Image.open(BytesIO(encoded_tensor))

CodePudding user response：

The error here is caused because you are not decoding the image. To decode the image, use tf.io.decode_jpeg. Please find the working code below.

import tensorflow as tf
from PIL import Image

img=tf.keras.utils.load_img('/content/dog.1.jpg')
img=tf.keras.utils.img_to_array(img)

#encode_jpeg encodes a tensor of type uint8 to string
encode_img=tf.io.encode_jpeg(img,'rgb')

#decode_jpeg decodes the string tensor to a tensor of type uint8
decode_img=tf.io.decode_jpeg(encode_img)

Image.fromarray(decode_img.numpy())