basicsr.utils.diffjpeg¶

Modified from https://github.com/mlomnitz/DiffJPEG

For images not divisible by 8 https://dsp.stackexchange.com/questions/35339/jpeg-dct-padding/35343#35343

class basicsr.utils.diffjpeg.BlockMerging[source]¶

Bases: Module

Merge patches into image

forward(patches, height, width)[source]¶

Parameters:

patches (tensor) –
height (int) –
width (int) –

Returns:

batch x height x width

Return type:

Tensor

training: bool¶

class basicsr.utils.diffjpeg.BlockSplitting[source]¶

Bases: Module

Splitting image into patches

forward(image)[source]¶

Parameters:: image (tensor) – batch x height x width
Returns:: batch x h*w/64 x h x w
Return type:: Tensor

training: bool¶

class basicsr.utils.diffjpeg.CDequantize[source]¶

Bases: Module

Dequantize CbCr channel

forward(image, factor=1)[source]¶

Parameters:: image (tensor) – batch x height x width
Returns:: batch x height x width
Return type:: Tensor

training: bool¶

class basicsr.utils.diffjpeg.CQuantize(rounding)[source]¶

Bases: Module

JPEG Quantization for CbCr channels

Parameters:: rounding (function) – rounding function to use

forward(image, factor=1)[source]¶

Parameters:: image (tensor) – batch x height x width
Returns:: batch x height x width
Return type:: Tensor

training: bool¶

class basicsr.utils.diffjpeg.ChromaSubsampling[source]¶

Bases: Module

Chroma subsampling on CbCr channels

forward(image)[source]¶

Parameters:: image (tensor) – batch x height x width x 3
Returns:: batch x height x width cb(tensor): batch x height/2 x width/2 cr(tensor): batch x height/2 x width/2
Return type:: y(tensor)

training: bool¶

class basicsr.utils.diffjpeg.ChromaUpsampling[source]¶

Bases: Module

Upsample chroma layers

forward(y, cb, cr)[source]¶

Parameters:

y (tensor) – y channel image
cb (tensor) – cb channel
cr (tensor) – cr channel

Returns:

batch x height x width x 3

Return type:

Tensor

training: bool¶

class basicsr.utils.diffjpeg.CompressJpeg(rounding=<built-in method round of type object>)[source]¶

Bases: Module

Full JPEG compression algorithm

Parameters:: rounding (function) – rounding function to use

forward(image, factor=1)[source]¶

Parameters:: image (tensor) – batch x 3 x height x width
Returns:: Compressed tensor with batch x h*w/64 x 8 x 8.
Return type:: dict(tensor)

training: bool¶

class basicsr.utils.diffjpeg.DCT8x8[source]¶

Bases: Module

Discrete Cosine Transformation

forward(image)[source]¶

Parameters:: image (tensor) – batch x height x width
Returns:: batch x height x width
Return type:: Tensor

training: bool¶

class basicsr.utils.diffjpeg.DeCompressJpeg(rounding=<built-in method round of type object>)[source]¶

Bases: Module

Full JPEG decompression algorithm

Parameters:: rounding (function) – rounding function to use

forward(y, cb, cr, imgh, imgw, factor=1)[source]¶

Parameters:

compressed (dict(tensor)) – batch x h*w/64 x 8 x 8
imgh (int) –
imgw (int) –
factor (float) –

Returns:

batch x 3 x height x width

Return type:

Tensor

training: bool¶

class basicsr.utils.diffjpeg.DiffJPEG(differentiable=True)[source]¶

Bases: Module

This JPEG algorithm result is slightly different from cv2. DiffJPEG supports batch processing.

Parameters:: differentiable (bool) – If True, uses custom differentiable rounding function, if False, uses standard torch.round

forward(x, quality)[source]¶

Parameters:

x (Tensor) – Input image, bchw, rgb, [0, 1]
quality (float) – Quality factor for jpeg compression scheme.

training: bool¶

class basicsr.utils.diffjpeg.RGB2YCbCrJpeg[source]¶

Bases: Module

Converts RGB image to YCbCr

forward(image)[source]¶

Parameters:: image (Tensor) – batch x 3 x height x width
Returns:: batch x height x width x 3
Return type:: Tensor

training: bool¶

class basicsr.utils.diffjpeg.YCbCr2RGBJpeg[source]¶

Bases: Module

Converts YCbCr image to RGB JPEG

forward(image)[source]¶

Parameters:: image (tensor) – batch x height x width x 3
Returns:: batch x 3 x height x width
Return type:: Tensor

training: bool¶

class basicsr.utils.diffjpeg.YDequantize[source]¶

Bases: Module

Dequantize Y channel

forward(image, factor=1)[source]¶

Parameters:: image (tensor) – batch x height x width
Returns:: batch x height x width
Return type:: Tensor

training: bool¶

class basicsr.utils.diffjpeg.YQuantize(rounding)[source]¶

Bases: Module

JPEG Quantization for Y channel

Parameters:: rounding (function) – rounding function to use

forward(image, factor=1)[source]¶

Parameters:: image (tensor) – batch x height x width
Returns:: batch x height x width
Return type:: Tensor

training: bool¶

basicsr.utils.diffjpeg.diff_round(x)[source]¶: Differentiable rounding function

class basicsr.utils.diffjpeg.iDCT8x8[source]¶

Bases: Module

Inverse discrete Cosine Transformation

forward(image)[source]¶

Parameters:: image (tensor) – batch x height x width
Returns:: batch x height x width
Return type:: Tensor

training: bool¶

basicsr.utils.diffjpeg.quality_to_factor(quality)[source]¶

Calculate factor corresponding to quality

Parameters:: quality (float) – Quality for jpeg compression.
Returns:: Compression factor.
Return type:: float