Welcome to compimg’s documentation!¶

compimg¶

PyPI PyPI - Python Version PyPI - Wheel License Documentation Status

Branches:master: CircleCIdevelop: CircleCI

Introduction¶

For full documentation visit documentation site.

Image similarity metrics are often used in image quality assessment for performance evaluation of image restoration and reconstruction algorithms. They require two images:

test image (image of interest)
reference image (image we compare against)

Such metrics produce numerical value.

Such methods are are widely called full/reduced-reference methods for assessing image quality.

compimg package is all about calculating similarity between images. It provides image similarity metrics (PSNR, SSIM etc.) that are widely used to asses image quality.

import numpy as np
from compimg.similarity import SSIM
some_grayscale_image = np.ones((20,20), dtype=np.uint8)
identical_image = np.ones((20,20), dtype=np.uint8)
result = SSIM().compare(some_grayscale_image, identical_image)
assert result == 1.0

Features¶

common metrics for calculating similarity of one image to another
only numpy as a dependency

Installation¶

compimg is available on PyPI. You can install it using pip:pip install compimg

Note¶

Keep in mind that metrics are not aware of what kind of image you are passing. If metric relies on intensity values and you have YCbCr image you should pass only the first channel to the computing routine.

Help¶

If you have any problems or questions please post an issue.

compimg package¶

How to use¶

Here is the simple example of how one can compare one image to another.

>>> import numpy as np
>>> from compimg.similarity import MSE
>>> img = np.ones((20,20), dtype = np.uint8)
>>> reference = np.ones((20,20), dtype = np.uint8)
>>> MSE().compare(img, img)
0.0

All metrics implement single interface so it is easy to use multiple of them for example you could run:

>>> import numpy as np
>>> from compimg.similarity import MSE, PSNR, SSIM
>>> for metric in [MSE(), PSNR(), SSIM()]:
...     img = np.ones((20,20), dtype = np.uint8)
...     reference = np.zeros((20,20), dtype = np.uint8)
...     value = round(metric.compare(img, reference), 2)
...     print(f"{metric.__class__.__name__} = {value}")
MSE = 1.0
PSNR = 48.13
SSIM = 0.87

compimg implicitly converts image to intermediate type (float64) to avoid overflow/underflow when doing calculation. Its advised to leave this type as is, albeit it is possible to change it. For example you could sacrafice precision to improve processing speed by changing it to float32 or even float16.

>>> import numpy as np
>>> import compimg
>>> import compimg.similarity
>>> compimg.config.intermediate_type = np.dtype(np.float32)
>>> # code that uses similarity metrics

Submodules¶

compimg.exceptions module¶

compimg exceptions module

exception compimg.exceptions.DifferentDTypesError(dtype1: numpy.dtype, dtype2: numpy.dtype)[source]¶: Bases: Exception

exception compimg.exceptions.DifferentShapesError(shape1: Sequence[int], shape2: Sequence[int])[source]¶: Bases: Exception

exception compimg.exceptions.KernelBiggerThanImageError(kernel_shape: Sequence[int], image_shape: Sequence[int])[source]¶: Bases: Exception

exception compimg.exceptions.KernelShapeNotOddError(kernel_shape: Sequence[int])[source]¶: Bases: Exception

exception compimg.exceptions.NegativePadAmountError(amount)[source]¶: Bases: Exception

compimg.similarity module¶

Module with routines for computing similarity between images

class compimg.similarity.GSSIM(k1: float = 0.01, k2: float = 0.03)[source]¶

Bases: compimg.similarity.SimilarityMetric

Gradient-Based Structural similarity index according to the paper “GRADIENT-BASED STRUCTURAL SIMILARITY FOR IMAGE QUALITY ASSESSMENT” by Chen et al.