CODEWITHSUNDEEP
pythonimagetensorflow
marko
marko

Reputation: 101

tf.image.decode_jpeg() returns matrix of zeros

I am using a script for loading images to TensorFlow which apparently works for everyone, but when I try it, I end up with black image (zero matrix). I tried it with several image files, and it's always zero, and when I deliberately misspell the image location string, it reports an error (as it should). Size of the returned image tensor is correct (256,256,3). This is the script, does someone sees the error? 

file_names = ['/home/marko/Data/train_27.jpg']
filename_queue = tf.train.string_input_producer(file_names)

image_reader = tf.WholeFileReader()
title, image_file = image_reader.read(filename_queue)

image = tf.image.decode_jpeg(image_file,channels=3)

with tf.Session() as sess:
    tf.global_variables_initializer().run()

    coord = tf.train.Coordinator()
    threads = tf.train.start_queue_runners(coord=coord)

    image_tensor = sess.run(image)
    print(image_tensor)
    print(image_tensor.shape)

    coord.request_stop()
    coord.join(threads)

Upvotes: 2

Views: 1252

Answers (2)

Evgeny Petrov
Evgeny Petrov

Reputation: 79

Your code is correct. I had the same problem with images from Kaggle competition

Seems like Tensorflow detects colorspace incorrectly for these files, or the colorspace information encoded in the images is incorrect.

It seems like Tensorflow doesn't allow to enforce colorspace while decoding images. So probably the easiest way is to "fix" the images.

I've used 'convert' utility from ImageMagic toolkit:

    ls train-jpg/ | \
        parallel convert -colorspace sRGB train-jpg/{} fixed-train-jpg/{}

Upvotes: 2

jerpint
jerpint

Reputation: 409

I have another solution that does not require conversion via command line as suggested by Evgeny, rather uses Pillow for loading the images:

import numpy as np
from PIL import Image

def load_image(img_path, resize=[256, 256]):

    pil_img = Image.open(img_path).convert("RGB")
    img = np.asarray(pil_img) / 255
    img_tensor = tf.convert_to_tensor(img)
    img_final = tf.image.resize(img_tensor, resize)
    return img_final

Upvotes: 1

Related Questions