dct

Batch block discrete cosine transform (DCT) methods.

Given a signal, the Fourier transform decomposes that signal into a basis representation based on its constituent frequencies. The DCT can be thought of as a special case of the discrete Fourier transform (DFT) for real-valued signals. The DCT (and more generally the DFT) can be computed in O(n log n) time, where n is the number of samples of the signal.

This module was first introduced to better performing DCT and inverse DCT operations on blocks of custom size, which is required for the frequency space perturbed patch cloak noise layer.

For more information on discrete cosine transforms see: https://en.wikipedia.org/wiki/Discrete_cosine_transform https://math.mit.edu/~gs/papers/dct.pdf

Functions:

Name	Description
blockwise_dct	Decompose batches of images into non-overlapping square blocks of length patch_size, performs a DCT-II on each block, and then
blockwise_idct	Decompose batches of images into non-overlapping square blocks of length patch_size, performs an inverse DCT-II on each block, and
calculate_block_avg_colors	Calculate the average pixel-space value of each image and channel per patch on a given tensor.
set_block_average_colors_on_tensor	Replace the (0, 0) mode of each input per image per patch per channel with the supplied block_avg_colors tensor.

blockwise_dct

blockwise_dct(input: Tensor, patch_size: tuple, image_size: tuple) -> <class 'torch.Tensor'>

Decompose batches of images into non-overlapping square blocks of length patch_size, performs a DCT-II on each block, and then reshape the blocks back into a single image-like tensor of shape image_size.

Notes

The transformations supplied by torchjpeg are orthogonal.

Parameters:

Name	Type	Description	Default
input	Tensor	The tensor to be transformed.	required
patch_size	tuple	The length of the square blocks to apply the DCT-II to.	required
image_size	tuple	The original to transform the non-overlapping blocks to.	required

Returns:

Type	Description
<class 'torch.Tensor'>	A tensor which has been block transformed by DCT-II.

blockwise_idct

blockwise_idct(input: Tensor, patch_size: tuple, image_size: tuple) -> <class 'torch.Tensor'>

Decompose batches of images into non-overlapping square blocks of length patch_size, performs an inverse DCT-II on each block, and then reshape the blocks back into a single image-like tensor of shape image_size.

Notes

The inverse of DCT-II is DCT-III. The transformations supplied by torchjpeg are orthogonal.

Parameters:

Name	Type	Description	Default
input	Tensor	The tensor to be transformed.	required
patch_size	tuple	The length of the square blocks to apply the inverse DCT-II to.	required
image_size	tuple	The tensor shape to transform the non-overlapping blocks to.	required

Returns:

Type	Description
<class 'torch.Tensor'>	A tensor which has been block transformed by inverse DCT-II.

calculate_block_avg_colors

calculate_block_avg_colors(input: Tensor, patch_size: tuple) -> <class 'torch.Tensor'>

Calculate the average pixel-space value of each image and channel per patch on a given tensor.

The (0, 0) mode of the DCT is the average. The input is assumed to be in the DCT representation.

Parameters:

Name	Type	Description	Default
input	Tensor	The tensor to get the (0, 0) DCT mode from.	required
patch_size	tuple	The (H, W) size of the patches.	required

Returns:

Type	Description
<class 'torch.Tensor'>	The resulting tensor of averages.

set_block_average_colors_on_tensor

set_block_average_colors_on_tensor(input: Tensor, block_avg_colors: Tensor, patch_size: tuple, image_size: tuple) -> <class 'torch.Tensor'>

Replace the (0, 0) mode of each input per image per patch per channel with the supplied block_avg_colors tensor.

The tensors are assumed to be in the DCT representation, and thus the (0, 0) mode represents a per image, per patch, per channel average value.

Parameters:

Name	Type	Description	Default
input	Tensor	The tensor to perform the replacement on.	required
block_avg_colors	Tensor	The tensor which represents the per image per patch per channel averages.	required
patch_size	tuple	The (H, W) size of the patches.	required
image_size	tuple	The (H, W) size of the image.	required

Returns:

Type	Description
<class 'torch.Tensor'>	The input tensor where the (0, 0) DCT modes have been replaced by block_avg_colors.