tf.keras.ops.image.extract_patches

Extracts patches from the image(s) or volume(s).

tf.keras.ops.image.extract_patches(
    images,
    size,
    strides=None,
    dilation_rate=1,
    padding='valid',
    data_format=None
)

This function supports both 2D and 3D patch extraction based on the size argument length, similar to how keras.ops.conv handles different dimensions.

images Input image/volume or batch of images/volumes. For 2D patches: 3D (H, W, C) or 4D (N, H, W, C). For 3D patches: 4D (D, H, W, C) or 5D (N, D, H, W, C).
size Patch size as int or tuple. Length 2 tuple (patch_height, patch_width) or int for 2D patches. Length 3 tuple (patch_depth, patch_height, patch_width) for 3D patches.
strides Strides for patch extraction. If not specified, defaults to size (non-overlapping patches).
dilation_rate Dilation rate for patch extraction. Note that dilation_rate > 1 is not supported with strides > 1.
padding The type of padding algorithm to use: "same" or "valid".
data_format A string specifying the data format of the input tensor. It can be either "channels_last" or "channels_first". If not specified, defaults to keras.config.image_data_format.

Extracted patches with shape depending on input and size:

  • 2D patches: 3D (unbatched) or 4D (batched)
  • 3D patches: 4D (unbatched) or 5D (batched)

Examples:

# 2D patches from batch of images
image = np.random.random(
    (2, 20, 20, 3)
).astype("float32")
patches = keras.ops.image.extract_patches(image, (5, 5))
patches.shape
(2, 4, 4, 75)
 
# 2D patches from single image
image = np.random.random((20, 20, 3)).astype("float32")
patches = keras.ops.image.extract_patches(image, (3, 3), (1, 1))
patches.shape
(18, 18, 27)
 
# 3D patches from batch of volumes
volumes = np.random.random(
    (2, 10, 10, 10, 3)
).astype("float32")
patches = keras.ops.image.extract_patches(volumes, (3, 3, 3))
patches.shape
(2, 3, 3, 3, 81)
 
# 3D patches from single volume
volume = np.random.random((10, 10, 10, 3)).astype("float32")
patches = keras.ops.image.extract_patches(volume, (3, 3, 3))
patches.shape
(3, 3, 3, 81)