net.imglib2.algorithm.metrics.imagequality

Class SSIM

java.lang.Object

net.imglib2.algorithm.metrics.imagequality.SSIM

public class SSIM
extends Object

Compute the structural similarity index measure (SSIM) between a reference and a processed image. The SSIM is implemented here by using Gaussian weighting when computing the mean, variance and covariance of small regions surrounding each pixel. The Gaussian filter standard deviation is set as a parameter. The original publication used sigma = 1.5. The k1 and k2 SSIM parameters are fixed, with K1 = 0.01 and K2 = 0.03.

The SSIM is run on 2D or 3D images. Because padding is applied to avoid edge effects in all dimensions, the minimum dimension size must be big enough to prevent negative dimensions. Depending on sigma, different algorithms are called to perform the Gaussian convolutions (Gauss3 or FastGauss). Which yields the following constraints:

sigma > 2: d > 2 * floor( 3.7210 * sigma + 0.20157 + 0.5 )

0.5 <= sigma <= 2: d > 2 * floor( 3 * sigma + 0.5)

sigma < 0.5: d > 4

For sigma = 1.5, the minimum dimension depth is then 11.

Note that the SSIM metrics throws an exception if the images have singular dimensions for the aforementioned reason.

Author:

Joran Deschamps

See Also:

Original publication, Wikipedia page

Field Summary

Fields

Modifier and Type	Field and Description
static double	K1
static double	K2

Constructor Summary

Constructors

Constructor and Description
SSIM()

Method Summary

Modifier and Type	Method and Description
static <T extends RealType<T>> double	computeMetrics(RandomAccessibleInterval<T> reference, RandomAccessibleInterval<T> processed, double sigma) Compute the SSIM metrics.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Detail

K1

public static final double K1

See Also:

Constant Field Values

K2

public static final double K2

See Also:

Constant Field Values

Constructor Detail

SSIM

public SSIM()

Method Detail

computeMetrics

public static <T extends RealType<T>> double computeMetrics(RandomAccessibleInterval<T> reference,
                                                            RandomAccessibleInterval<T> processed,
                                                            double sigma)

Compute the SSIM metrics. The original publications used sigma = 1.5. The method expects 2D or 3D images, no singular dimensions, and a minimum depth for each dimension that allows padding on both sides.

Type Parameters:

T - Pixel type of the images

Parameters:

reference - Reference image

processed - Processed image

sigma - Standard deviation of the Gaussian filter

Returns:

SSIM metrics score