tf.keras.ops.image.reconstruct_patches

Reconstructs image(s) or volume(s) from non-overlapping patches.

tf.keras.ops.image.reconstruct_patches(
    patches,
    size,
    output_size=None,
    strides=None,
    padding='valid',
    data_format=None
)

Inverse of keras.ops.image.extract_patches for the non-overlapping case (strides == size).

patches Patches tensor as produced by extract_patches. For 2D patches: 3D (gH, gW, pH*pW*C) or 4D (N, gH, gW, pH*pW*C). For 3D patches: 4D (gD, gH, gW, pD*pH*pW*C) or 5D (N, gD, gH, gW, pD*pH*pW*C). With data_format="channels_first" the flat patch dim comes first instead: (pH*pW*C, gH, gW) / (N, pH*pW*C, gH, gW) for 2D patches, (pD*pH*pW*C, gD, gH, gW) / (N, pD*pH*pW*C, gD, gH, gW) for 3D patches.
size Patch size, matching the size used for extraction. Length 2 tuple for 2D, length 3 tuple for 3D, or int.
output_size Target spatial shape of the reconstruction. Length 2 tuple (H, W) for 2D, length 3 tuple (D, H, W) for 3D. With padding="valid" this may be omitted (None) and is then inferred from the patch grid; if given, it must equal (grid - 1) * stride + size per dim — the region covered by the extracted patches, i.e. the original size cropped down to a multiple of size in the non-overlapping case. With padding="same" it is required: pass the original spatial shape so the padding added during extraction can be unambiguously removed.
strides Currently must equal size (non-overlapping). Defaults to size.
padding "same" or "valid", matching the extraction.
data_format A string specifying the data format of the input tensor. Either "channels_last" or "channels_first". With "channels_first" the reconstruction is returned in (C, H, W) / (N, C, H, W) layout. If not specified, defaults to keras.config.image_data_format.

Reconstructed image/volume: 2D: 3D (unbatched) or 4D (batched). 3D: 4D (unbatched) or 5D (batched).

Examples:

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)
recon = keras.ops.image.reconstruct_patches(
    patches, size=(5, 5), output_size=(20, 20)
)
recon.shape
(2, 20, 20, 3)
 
# 3D patches: pass a length-3 `size`
volume = np.random.random((2, 9, 9, 9, 3)).astype("float32")
patches = keras.ops.image.extract_patches(volume, (3, 3, 3))
recon = keras.ops.image.reconstruct_patches(
    patches, size=(3, 3, 3), output_size=(9, 9, 9)
)
recon.shape
(2, 9, 9, 9, 3)