/TensorFlow 2.4


Resize images to size using the specified method.

Resized images will be distorted if their original aspect ratio is not the same as size. To avoid distortions see tf.image.resize_with_pad.

image = tf.constant([
# Add "batch" and "channels" dimensions
image = image[tf.newaxis, ..., tf.newaxis]
image.shape.as_list()  # [batch, height, width, channels]
[1, 5, 5, 1]
tf.image.resize(image, [3,5])[0,...,0].numpy()
array([[0.6666667, 0.3333333, 0.       , 0.       , 0.       ],
       [0.       , 0.       , 1.       , 0.       , 0.       ],
       [0.       , 0.       , 0.       , 0.3333335, 0.6666665]],

It works equally well with a single image instead of a batch of images:

tf.image.resize(image[0], [3,5]).shape.as_list()
[3, 5, 1]

When antialias is true, the sampling filter will anti-alias the input image as well as interpolate. When downsampling an image with anti-aliasing the sampling filter kernel is scaled in order to properly anti-alias the input image signal. antialias has no effect when upsampling an image:

a = tf.image.resize(image, [5,10])
b = tf.image.resize(image, [5,10], antialias=True)
tf.reduce_max(abs(a - b)).numpy()

The method argument expects an item from the image.ResizeMethod enum, or the string equivalent. The options are:

  • bilinear: Bilinear interpolation. If antialias is true, becomes a hat/tent filter function with radius 1 when downsampling.
  • lanczos3: Lanczos kernel with radius 3. High-quality practical filter but may have some ringing, especially on synthetic images.
  • lanczos5: Lanczos kernel with radius 5. Very-high-quality filter but may have stronger ringing.
  • bicubic: Cubic interpolant of Keys. Equivalent to Catmull-Rom kernel. Reasonably good quality and faster than Lanczos3Kernel, particularly when upsampling.
  • gaussian: Gaussian kernel with radius 3, sigma = 1.5 / 3.0.
  • nearest: Nearest neighbor interpolation. antialias has no effect when used with nearest neighbor interpolation.
  • area: Anti-aliased resampling with area interpolation. antialias has no effect when used with area interpolation; it always anti-aliases.
  • mitchellcubic: Mitchell-Netravali Cubic non-interpolating filter. For synthetic images (especially those lacking proper prefiltering), less ringing than Keys cubic kernel but less sharp.
Note: Near image edges the filtering kernel may be partially outside the image boundaries. For these pixels, only input pixels inside the image will be included in the filter sum, and the output value will be appropriately normalized.

The return value has type float32, unless the method is ResizeMethod.NEAREST_NEIGHBOR, then the return dtype is the dtype of images:

nn = tf.image.resize(image, [5,7], method='nearest')
array([[1, 0, 0, 0, 0, 0, 0],
       [0, 1, 1, 0, 0, 0, 0],
       [0, 0, 0, 1, 0, 0, 0],
       [0, 0, 0, 0, 1, 1, 0],
       [0, 0, 0, 0, 0, 0, 1]], dtype=int32)

With preserve_aspect_ratio=True, the aspect ratio is preserved, so size is the maximum for each dimension:

max_10_20 = tf.image.resize(image, [10,20], preserve_aspect_ratio=True)
[1, 10, 10, 1]
images 4-D Tensor of shape [batch, height, width, channels] or 3-D Tensor of shape [height, width, channels].
size A 1-D int32 Tensor of 2 elements: new_height, new_width. The new size for the images.
method An image.ResizeMethod, or string equivalent. Defaults to bilinear.
preserve_aspect_ratio Whether to preserve the aspect ratio. If this is set, then images will be resized to a size that fits in size while preserving the aspect ratio of the original image. Scales up the image if size is bigger than the current size of the image. Defaults to False.
antialias Whether to use an anti-aliasing filter when downsampling an image.
name A name for this operation (optional).
ValueError if the shape of images is incompatible with the shape arguments to this function
ValueError if size has an invalid shape or type.
ValueError if an unsupported resize method is specified.
If images was 4-D, a 4-D float Tensor of shape [batch, new_height, new_width, channels]. If images was 3-D, a 3-D float Tensor of shape [new_height, new_width, channels].

© 2020 The TensorFlow Authors. All rights reserved.
Licensed under the Creative Commons Attribution License 3.0.
Code samples licensed under the Apache 2.0 License.