Tensorflow无法正确解码图像

Question

I am new to tensorflow. I am reading images from files and decoding them with tf.image.decode_jpeg and then I am plotting decoded image with matplotlib. But somehow original and decoded images are different.

filenames = ['/Users/darshak/TensorFlow/100.jpg', '/Users/darshak/TensorFlow/10.jpg']
filename_queue = tf.train.string_input_producer(filenames)

reader = tf.WholeFileReader()
filename, content = reader.read(filename_queue)

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

image = tf.cast(image, tf.float32)

resized_image = tf.image.resize_images(image, [256, 256])

image_batch = tf.train.batch([resized_image], batch_size=9)

sess = tf.InteractiveSession()

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

plt.imshow(image.eval())
plt.show()
sess.close()

Answer 1

The problem arises because plt.imshow(image.eval()) interprets the image data different depending on the element type of image .

• If image is a tf.uint8 tensor (ie as it is produced by tf.image.decode_jpeg() ) it will contain values from 0 to 255 for the R, G, and B channels, and plt.imshow() interprets (0, 0, 0) as black and (255, 255, 255) as white.

• When you cast image to be a tf.float32 tensor, it will contain values from 0.0 to 255.0 for the R, G, and B channels, and plt.imshow() interprets (0.0, 0.0, 0.0) as black, but it interprets (1.0, 1.0, 1.0) as white. All values greater than 1.0 are treated the same as 1.0 , and as a result the image appears discolored.

If you intend to represent the image as a tf.float32 tensor and visualize it, you should divide the image values by 255.0 :

image = tf.image.decode_jpeg(content, channels=3)
image = tf.cast(image, tf.float32) / 255.0