Different results from scipy imread and tensorflow decode_jpeg

[HaaPut]

Problem Description

Image decoded using decode_jpeg from tensor flow is visually similar, but numerically different from one returned by scipy imread.

Minimal Example:

import numpy as np
import scipy
import tensorflow as tf
def minimal_example():
    #image_source = 'https://upload.wikimedia.org/wikipedia/commons/8/88/Astronaut-EVA.jpg'
    image_path = 'astronaut.jpg'
    image_file = open(image_path,'rb')
    image_raw = image_file.read()
    image_scipy = scipy.misc.imread(image_path)
    image_tf = tf.image.decode_jpeg(image_raw).eval(session=tf.Session())
    print('Error: ', np.sum(np.abs(image_tf - image_scipy)))
    #Error:  3420883624

System information

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow): Yes
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Linux Ubuntu 16.04
• TensorFlow installed from (source or binary): binary
• TensorFlow version (use command below): v1.2.0-rc2-21-g12f033d 1.2.0
• Python version: 3.5

[yangda75]

I can reproduce the result with exactly the same error using the same image and code.
Can this be caused by decoders' differences? Since scipy.misc.imread uses pillow, tf.image.decode_jpeg still functions without pillow.

system information

• OS: Fedora 26
• Python version: 3.5.3
• Python related management tool: Anaconda 3
• Tensorflow(installed via conda install) version: 1.1.0, np112py35_0

printed a small portion of the results

[1:10,:,2] scipy image
 [[0 0 0 ..., 0 2 1]
 [0 0 0 ..., 0 5 4]
 [0 0 0 ..., 0 5 4]
 ..., 
 [0 0 2 ..., 0 5 5]
 [0 0 0 ..., 0 3 0]
 [0 0 0 ..., 3 3 0]]
[1:10,:,2] tf image
 [[0 0 0 ..., 0 0 0]
 [0 0 0 ..., 0 3 4]
 [0 0 0 ..., 0 3 4]
 ..., 
 [0 0 0 ..., 0 3 3]
 [0 0 0 ..., 0 2 0]
 [0 0 0 ..., 0 2 0]]

[drpngx]

We use libjpeg. I am not sure what your application is. If you really need this, I would recommend writing your own op on top of pillow.

I am going to close this, since I don't think it warrants changing the TF core.

[HaaPut]

I think the reason for this difference is that Tensorflow used fast integer DCT as a default decode jpeg image

tensorflow/tensorflow/core/kernels/decode_image_op.cc

Line 109 in e073322

// The TensorFlow-chosen default for jpeg decoding is IFAST, sacrificing

[drpngx]

Oh yes, I remember some discussion about that. I think we change the default but the other version is available through options.