returns a saliency map for image.

Details and Options

  • Saliency filtering produces a single-channel image (also known as the saliency map) in which intensities represent how important or noticeable the pixel is in the input image.
  • ImageSaliencyFilter is typically used to focus on regions of an image that have features that stand out as different. A higher saliency value is taken to be more important.
  • ImageSaliencyFilter works on 2D images with an arbitrary number of channels.
  • ImageSaliencyFilter takes a Method option. Possible settings are:
  • "Itti"saliency based on color, intensity and orientation
    "IttiColor"color map of the Itti algorithm
    "IttiIntensity"intensity map of the Itti algorithm
    "IttiOrientation"orientation map of the Itti algorithm
    "HistogramContrast"saliency based on mean color difference to image pixels
    "SpectralResidual"saliency based on the residual of the log spectrum
    "ImageSignature"saliency based on foreground estimation
    "U2Net"U^2-Net salient object detection (default)
  • The Itti algorithm is based on human visual saliency. For all methods based on this algorithm, calculations are performed in the RGB color space.
  • For all Itti-based methods, different scales of the image are compared with each other. With Method->{"method",{σ1,σ2,},{step1,step2,}}, scale σi is compared with coarser scales σi+stepj for all j. The default setting is {"Itti",{2,3,4},{3,4}}.
  • Using Method{"U2Net",TargetDevicedevice}, the device on which computation is performed can be specified.


open allclose all

Basic Examples  (2)

Saliency filtering of a grayscale image:

Saliency filtering of a color image:

Options  (6)

Method  (6)

By default, "U2Net" is used:

Compute a saliency based on the Itti visual attention model:

Return orientation and intensity maps used by the Itti algorithm:

The "HistogramContrast" method estimates the color contrast using a quantized image:

Specify the number of levels to be used:

Compute the saliency with the "ImageSignature" method:

Compute the saliency with the "SpectralResidual" method:

Saliency filtering using the "U2Net" neural network:

Applications  (4)

Use saliency to separate the image foreground from its background:

Compute the image saliency map:

Extract the foreground region by binarizing the saliency map using an extended maximum:

Create a mask identifying the most salient objects in a scene:

Binarize the saliency map and keep the biggest components:

Extract the salient regions from the image:

Trim the salient regions of the image:

Extract the salient region of an image using the saliency map:

Trim the region with max saliency:

Compute the saliency in different scales using image pyramids:

Compute the pyramid, apply saliency filtering at all levels:

Reconstruct the multiscale saliency map by accumulating the result at all levels:

Show the map on the images as the alpha channel:

Compare to the single-scale saliency filtering:

Compare to neural networkbased saliency filtering:

Properties & Relations  (1)

With Method->"Itti", the result is computed faster when fewer comparison maps are used:

Itti is still faster than the neural networkbased method:

Wolfram Research (2014), ImageSaliencyFilter, Wolfram Language function, (updated 2023).


Wolfram Research (2014), ImageSaliencyFilter, Wolfram Language function, (updated 2023).


Wolfram Language. 2014. "ImageSaliencyFilter." Wolfram Language & System Documentation Center. Wolfram Research. Last Modified 2023.


Wolfram Language. (2014). ImageSaliencyFilter. Wolfram Language & System Documentation Center. Retrieved from


@misc{reference.wolfram_2024_imagesaliencyfilter, author="Wolfram Research", title="{ImageSaliencyFilter}", year="2023", howpublished="\url{}", note=[Accessed: 26-May-2024 ]}


@online{reference.wolfram_2024_imagesaliencyfilter, organization={Wolfram Research}, title={ImageSaliencyFilter}, year={2023}, url={}, note=[Accessed: 26-May-2024 ]}