This project contains differentiable perceptual quality metrics implemented in the TensorFlow framework.

pip install https://github.com/google-research/perceptual-quality/archive/master.zip

This is an implementation of the perceptual information metric, as described in:

"An Unsupervised Information-Theoretic Perceptual Quality Metric"
S. Bhardwaj, I. Fischer, J. Ballé, T. Chinen
https://proceedings.neurips.cc/paper/2020/file/00482b9bed15a272730fcb590ffebddd-Paper.pdf

Usage:

from perceptual_quality import pim

model = pim.load_trained("pim-5")

# image_A, image_B: 4D tensors, batch x height x width x 3, sRGB colorspace.
# Returns the PIM distance between A and B for each batch element.
distance = model(image_A/255, image_B/255)

# Returns a `Distribution` object with the latent representation of the image.
distribution = model(image_A/255)

Refer to the online help of pim.PIM.call() for further information on supported image formats, etc.

This is an implementation of the normalized Laplacian pyramid distance, as described in:

"Perceptually optimized image rendering"
V. Laparra, A. Berardino, J. Ballé and E. P. Simoncelli
https://doi.org/10.1364/JOSAA.34.001511

Usage:

from perceptual_quality import nlpd

# image_A, image_B: at least 3D tensors, height x width x 3, sRGB or grayscale.
# Will return one number per color channel. Can also be batched.
distance = nlpd.nlpd(image_A, image_B)
distance = nlpd.nlpd_fast(image_A, image_B)

# Returns a list of tensors with the subband representation of the image.
model = nlpd.NLP()
subbands = model(image_A)

Refer to the online help of nlpd.nlpd(), nlpd.nlpd_fast(), and nlpd.NLP() for further information on supported image formats, assumed display characteristics, etc.

• Sangnie Bhardwaj (github: sangnie)
• Johannes Ballé (github: jonycgn)
• Ian Fischer (github: iansf)

Note that this is not an officially supported Google product.