itti1998 Code Reading

I think the core of the algorithm is to use the original image minus the Gaussian filter results to indicate the surrounds to the center of the inhibition, with different feature channels to express the different feelings of the eye nerves; n (.) emphasis on peaks;

1) Use default parameters

2) Calculate Gabor Filter

Select the Gabor angle from 0º~180º and other intervals for Param.gabor

Calculates the length and width of the saliency map, and the proportions of the original image.

Imgs{1, 2, 3} Three Center scale

If the original image does not have three channels (RGB) then feature maps does not have C and D channels

3) Calculate D-Channel feature Map:getchan ()

For each center scale (default of 3 128,64,32)

D-Channel:

If the original image has an RGB three color channel, the original image is converted from RGB to DKL space, three channels of dkl space are added to chan{}

Getchan () calculates the feature of the D-Channel when Maps returns, Chans contains 9 cells, respectively, the DKL component of the three scale.

4) Calculate the pixel saliency:pixsal () function for each feature map of the D channel

Call Mygausskernel () to generate a Gaussian kernel function for a given surrounds Sigma (default 2 and 8), one-dimensional

This feature map (local variable img) and Gaussian kernel vector ker do convolution, in and ker transpose to make convolution. This is Gaussian blur (not like this??? ）

The square of the difference between the original and the Gaussian blur is then reduced to 0, 1 (the result of the direct minus Gaussian filter)? In the article is cross scale minus ah???? ）

Call Mypeakiness () to do the N (.) described in the text Processing, which calls a mexlocalmaximagbvs to find the local extremum, is something in an open source library?

Finally add up the results of different surrounds Sigma to this and in the 0, 1

Taking 128*128 as an example, the D component and the respective saliency are:

For each maps of the D channel, all resize to mapsize, and then N (.). Weighted weights for the obtained weights and saliency map as the D-Channel

5) The mean value of three color components is directly obtained, which is the feature of I channel. I-channel 128*128 feature and Saliency are as follows

The weighted sum of the I channels is

6) O-Channel, 0º,128*128 feature and saliency:

O-Channel weighting and:

Three channels weighted sum: blurring processing (Gaussian filter blur): Center offset (multiplied by Gaussian function):

Finally to the final saliency map of 0, 1 intervals

itti1998 Code Reading