tf.keras.ops.image.ssim

Computes the Structural Similarity Index (SSIM) between two images.

tf.keras.ops.image.ssim(
    image1,
    image2,
    max_val=1.0,
    filter_size=11,
    filter_sigma=1.5,
    k1=0.01,
    k2=0.03,
    data_format=None
)

The SSIM index is a method for measuring the similarity between two images. It is based on the comparison of luminance, contrast, and structure between the images. The resulting SSIM index is a value between -1 and 1, where 1 indicates identical images.

This implementation is based on the original SSIM paper: Wang, Z., Bovik, A. C., Sheikh, H. R., & Simoncelli, E. P. (2004). "Image quality assessment: from error visibility to structural similarity." IEEE Transactions on Image Processing.

image1 First image or batch of images. Must be 3D or 4D.
image2 Second image or batch of images. Must have the same shape as image1.
max_val The maximum possible pixel value of the images. Defaults to 1.0 (for normalized images). Use 255.0 for images with pixel values in [0, 255].
filter_size Size of the Gaussian filter used for computing SSIM. Defaults to 11. Must be an odd integer >= 1.
filter_sigma Standard deviation of the Gaussian filter. Defaults to 1.5.
k1 First stabilization constant. Defaults to 0.01.
k2 Second stabilization constant. Defaults to 0.03.
data_format A string specifying the data format of the input tensor. It can be either "channels_last" or "channels_first". "channels_last" corresponds to inputs with shape (batch, height, width, channels), while "channels_first" corresponds to inputs with shape (batch, channels, height, width). If not specified, the value will default to keras.config.image_data_format.

A tensor of SSIM values. For batched inputs, returns a 1D tensor with one SSIM value per image pair. For unbatched inputs, returns a scalar tensor.

Examples:

import numpy as np
from keras import ops
# Two identical images should have SSIM = 1
image = np.random.random((32, 32, 3)).astype("float32")
ssim_value = ops.image.ssim(image, image, max_val=1.0)
float(ssim_value) > 0.99
True
 
# Batched images
images1 = np.random.random((2, 32, 32, 3)).astype("float32")
images2 = np.random.random((2, 32, 32, 3)).astype("float32")
ssim_values = ops.image.ssim(images1, images2, max_val=1.0)
ssim_values.shape
(2,)