Note: Functions taking
Tensorarguments can also take anything accepted by
TensorFlow provides Ops to decode and encode JPEG and PNG formats. Encoded images are represented by scalar string Tensors, decoded images by 3-D uint8 tensors of shape
[height, width, channels]. (PNG also supports uint16.)
The encode and decode Ops apply to one image at a time. Their input and output are all of variable size. If you need fixed size images, pass the output of the decode Ops to one of the cropping and resizing Ops.
Note: The PNG encode and decode Ops support RGBA, but the conversions Ops presently only support RGB, HSV, and GrayScale. Presently, the alpha channel has to be stripped from the image and re-attached using slicing ops.
The resizing Ops accept input images as tensors of several types. They always output resized images as float32 tensors.
The convenience function
tf.image.resize_images supports both 4-D and 3-D tensors as input and output. 4-D tensors are for batches of images, 3-D tensors for individual images.
# Decode a JPG image and resize it to 299 by 299 using default method. image = tf.image.decode_jpeg(...) resized_image = tf.image.resize_images(image, [299, 299])
Image ops work either on individual images or on batches of images, depending on the shape of their input Tensor.
If 3-D, the shape is
[height, width, channels], and the Tensor represents one image. If 4-D, the shape is
[batch_size, height, width, channels], and the Tensor represents
channels can usefully be 1, 2, 3, or 4. Single-channel images are grayscale, images with 3 channels are encoded as either RGB or HSV. Images with 2 or 4 channels include an alpha channel, which has to be stripped from the image before passing the image to most image processing functions (and can be re-attached later).
Internally, images are either stored in as one
float32 per channel per pixel (implicitly, values are assumed to lie in
[0,1)) or one
uint8 per channel per pixel (values are assumed to lie in
TensorFlow can convert between images in RGB or HSV. The conversion functions work only on float images, so you need to convert images in other formats using
# Decode an image and convert it to HSV. rgb_image = tf.image.decode_png(..., channels=3) rgb_image_float = tf.image.convert_image_dtype(rgb_image, tf.float32) hsv_image = tf.image.rgb_to_hsv(rgb_image)
TensorFlow provides functions to adjust images in various ways: brightness, contrast, hue, and saturation. Each adjustment can be done with predefined parameters or with random parameters picked from predefined intervals. Random adjustments are often useful to expand a training set and reduce overfitting.
If several adjustments are chained it is advisable to minimize the number of redundant conversions by first converting the images to the most natural data type and representation (RGB or HSV).
© 2017 The TensorFlow Authors. All rights reserved.
Licensed under the Creative Commons Attribution License 3.0.
Code samples licensed under the Apache 2.0 License.