PSNR-HVS-M, PSNR-HA and PSNR-HMA metrics for NumPy and PyTorch

Accelerated Python package for computing several image metrics based on human perception with backends in NumPy, PyTorch and C++.

This is an implementation of the PSNR-HVS, PSNR-HVS-M, PSNR-HA and PSNR-HMA metrics developed by Nikolay Ponomarenko.

The values produced by this library have been cross-checked against the results within the TID2013 dataset. (See the folder tid2013_results.) The only difference is that this library follows the common convention that PSNR for identical signals equals 100.0.

A miniscule discrepancy for PSNR-HMA (<0.01dB on average) is under investigation.

Bibliography

• Egiazarian, Karen, et al. "New full-reference quality metrics based on HVS." Proceedings of the Second International Workshop on Video Processing and Quality Metrics. Vol. 4. 2006.
• Ponomarenko, Nikolay, et al. "On between-coefficient contrast masking of DCT basis functions." Proceedings of the third international workshop on video processing and quality metrics. Vol. 4. 2007.
• Ponomarenko, Nikolay, et al. "Modified image visual quality metrics for contrast change and mean shift accounting." 2011 11th International Conference The Experience of Designing and Application of CAD Systems in Microelectronics (CADSM). IEEE, 2011.

Installation

psnr_hvsm supports Python 3.10-3.15. Packages are distributed on PyPi. Be sure to have an up-to-date pip to be able to install the correct packages on Linux:

python -m pip install --upgrade pip

pip install psnr_hvsm

With PyTorch support:

pip install psnr_hvsm[torch]

NumPy and PyTorch support policy

The NumPy requirement carries no upper bound. The compiled extension is built against NumPy 2 headers with a fixed backward-compatible ABI target, so a single wheel serves both NumPy 1.x and NumPy 2.x runtimes. Installing the package never forces a downgrade from NumPy 2.x, and its runtime floor is numpy>=1.19. Note that NumPy 1.x itself is only installable on Python 3.10-3.12: the last 1.x release ships wheels only that far, so on Python 3.13 and newer the runtime resolves NumPy 2.x.

PyPI publishes PyTorch wheels for Python 3.10-3.14 only, so the torch extra is unavailable on 3.15 and the PyTorch backend cannot be installed there.

MSVC runtime on Windows

Linux manylinux wheels bundle libstdc++ via auditwheel. The Windows wheels do not bundle the MSVC C++ runtime: CMake installs only the FFTW DLLs into the wheel. On CPython 3.10-3.14 the package therefore declares msvc-runtime for Windows, which pip installs into the environment along with the package, so no system-wide Visual C++ runtime is required. The MSVC or clang toolchain supplies the CRT when the extension is compiled, which is why it is not a build requirement.

On CPython 3.15 the requirement is not declared. msvc-runtime publishes its wheels per CPython ABI and has no cp315 wheel, so asking for it there would leave the newest supported interpreter uninstallable. A CPython install from python.org brings the C runtime it was built with (its installer updates the C Runtime Library on the machine when one is needed), so 3.15 works without anything extra. If an import nevertheless fails with a missing vcruntime140.dll or msvcp140.dll - a possibility with the minimal layouts such as the embeddable package, MSYS2 or NuGet-based installs - install the Microsoft Visual C++ 2015-2022 redistributable for your architecture, or use CPython 3.10-3.14, where the requirement is declared.

Usage

Command line

Command line support is an extra that pulls imageio:

pip install psnr_hvsm[command_line]

python -m psnr_hvsm original.png distorted.png

Choosing a backend

The backend can be set by setting the PSNR_HVSM_BACKEND environment variable. Valid backends are:

• numpy - pure NumPy using scipy.fft for DCT
• cpp - C++ using FFTW3
• torch - PyTorch; install as psnr_hvsm[torch] to install PyTorch as well

export PSNR_HVSM_BACKEND=torch
python -m psnr_hvsm original.png distorted.png

The default device for PyTorch is cuda but it can be changed by setting the PSNR_HVSM_TORCH_DEVICE environment variable.

As a library

The function psnr_hvs_hvsm accepts images as single-channel floating-point NumPy arrays. The images need to be normalised, i.e. the values need to be in the range [0,1]. This can be achieved by converting the image to float and dividing by the maximum value given the bit depth. For 8 bits per component this is 255.

The images must be padded to a multiple of 8 in each dimension.

from imageio import imread
from psnr_hvsm import psnr_hvs_hvsm, bt601ycbcr

image1 = imread('tests/baboon.png').astype(float) / 255
image2 = imread('tests/baboon_msk.png').astype(float) / 255

image1_y, *_ = bt601ycbcr(image1)
image2_y, *_ = bt601ycbcr(image2)

psnr_hvs, psnr_hvsm = psnr_hvs_hvsm(image1, image2)

print(psnr_hvs, psnr_hvsm)

34.427054505764424 51.64722121999962

If you need to measure PSNR-HVS and PSNR-HVS-M on an RGB image, you need to convert it to an YUV colorspace and pass in only the luma component.

BT.601 conversion precision

When 8-bit RGB input is converted to the BT.601 YCbCr colorspace, the luma and both chroma components are by default rounded to 8-bit code values. This quantisation matches the 8-bit YCbCr pipeline of the reference implementation and is what reproduces the published TID2013 reference values this library is cross-checked against, which is why it remains the default. Every metric output computed with the default is bit-identical to previous releases.

If you are porting the metrics and want to compare against a non-quantised reference, pass precision='float' to compute the same BT.601 matrix arithmetic without the rounding step:

from psnr_hvsm import bt601ycbcr
from psnr_hvsm.compute_all_metrics import compute_all_metrics

image1_y, image1_cb, image1_cr = bt601ycbcr(image1_uint8, precision='float')

psnr_hvs, psnr_hvsm, psnr_ha, psnr_hma, psnr = compute_all_metrics(
    image1_uint8, image2_uint8, precision='float')

The precision option ('8-bit' default, 'float') is accepted by bt601ycbcr in both the NumPy and PyTorch backends with identical semantics, and by the compute_all_metrics entry point, which threads it through every metric that depends on the quantised 8-bit BT.601 YCbCr conversion. It applies only when that conversion runs. For grayscale input or non-uint8 color input, it has no effect.

On the command line the same choice is exposed as --precision {8-bit,float} on python -m psnr_hvsm, and the JSON output (--json) records the requested precision, so batch results are self-describing. The requested precision is applied wherever the quantised conversion runs; for grayscale or non-uint8 color input, the field has no effect.

PyTorch support

The PyTorch backend can be used for use with gradient descent algorithms and computation on GPUs. In order to use the PyTorch backend, either install the package with the torch extra:

pip install psnr_hvsm[torch]

If your PyTorch installation was manual, you need torch-dct in order to use the PyTorch backend:

pip install "torch-dct>=0.1.6"

An important distinction is that the functions that expect 3-channel input now expect (...,C,H,W) format in the PyTorch implementation. The PyTorch backend can be enabled by importing it directly from psnr_hvsm.torch:

import torch
from imageio import imread
from psnr_hvsm.torch import psnr_hvs_hvsm, bt601ycbcr

image1 = imread('tests/baboon.png').astype(float) / 255
image2 = imread('tests/baboon_msk.png').astype(float) / 255

image1 = torch.tensor(image1, device='cuda').moveaxis(-1, -3)  # convert to (N,C,H,W) format
image2 = torch.tensor(image2, device='cuda').moveaxis(-1, -3)  # convert to (N,C,H,W) format

image1_y, *_ = bt601ycbcr(image1)
image2_y, *_ = bt601ycbcr(image2)

psnr_hvs, psnr_hvsm = psnr_hvs_hvsm(image1_y, image2_y)

Alternatively, set the PSNR_HVSM_BACKEND environment variable to torch:

import os
os.environ['PSNR_HVSM_BACKEND'] = 'torch'

from psnr_hvsm import psnr_hvs_hvsm
# rest of code
# ...

Note on gradients

Some operations in the calculation of the HVS-M MSE lead to problems with the gradient and therefore a parameter called masking_epsilon (defaulting to 0.0) has been added to the PyTorch versions of hvs_hvsm_mse_tiles, hvs_hvsm_mse and psnr_hvs_hvsm. Set it to a small value (determined by your own experimentation) if you need the result of HVS-M for gradient descent.

Computing metrics for the TID2013 dataset

If you have a copy of the TID2013 dataset, you can re-verify the metrics for yourself:

python -m psnr_hvsm.tid2013_metrics D:\tid2013\ .\tid2013_results\

Other exported functions

• hvs_hvsm_mse_tiles - compute HVS and HVS-M scores on all 8x8 tiles in the images, returns an array of numbers
• hvs_hvsm_mse - compute average HVS and HVS-M scores

Building

Dependencies

psnr_hvsm has several dependencies:

• FFTW3 >= 3.3.9
• pybind11
• xtensor
• xtensor-python

FFTW3 is always fetched by CMake at the pinned release listed in psnr_hvsm/CMakeLists.txt. The remaining, header-only dependencies are installed at pinned, mutually compatible versions when they are not already installed, so a C++ compiler, a current CMake and pip or uv are enough to build from source:

pip install --upgrade -r requirements-build.txt
pip wheel --no-deps -w dist .

If you prefer a conda environment, create the provided one and CMake will discover the dependencies from it instead of fetching them:

conda env create -f psnr_hvsm-dev.yml

Development mode

To install in development mode:

pip install --upgrade -r requirements.txt

Creating Python wheel

pip install --upgrade -r requirements-build.txt
python -m build --wheel

Verifying a clean source build

To build from source in a scratch virtualenv that has no conda environment for CMake to discover and to run the test suite against the result:

./clean-build-check.sh

Running tests on different versions of Python using tox

pip install --upgrade -r requirements-tox.txt
tox --parallel auto