Saliency region selection in large aerial imagery using multi-scale SLIC Segmentation

1: saliency region selection in large aerial imagery using multi-scale SLIC segmentation,Proc. Spie8360, airborne intelligence, surveillance, reconnaissance (ISR) systems and applications IX, 2012

2:

Several points worth noting in this paper:

(1)

At a glance, salient regions are stand out while the rest of the scene is neglected since they do not attract visual attention.

(2)

When large and mostly uniform background regions, such as land, sea and snow field, are segmented by the superpixels of smaller scale, most superpixels will be of the average size and regular shape, as shown in left figure.

(3)

With the presence of the objects of interest, the superpixels will adhere tightly to the object structural boundaries, resulting in varous sizes and stronugly deformed shapes, as shown in right figure.

(4)

The irregularity in size and shape of the superpixels may serve as a measure of saliency. The method is valid when the superpixel scale is smaller that of the regions of interest.

(5)

In the case when the sizes of the salient regions are unknown, or the objects of interest have varous sizes over the scene, the multi-scale SLIC may be applied with a set of the preset scale values.

(6)

Firstly, when the scale S is much smaller than the size of the salient structures, most superpixels will be of average size and regular shape.

Secondly, when the scale S is much larger than the size of the objects of interest, the object detail structures will be ignored in the SLIC segmentation.

(7)

In the multi-scale SLIC segmentation, the variation in size and shape of a superpixel associated to the presence of the object of interest can be characterized byHausdorff distanceDefined in the superpixel string-to-string presentation.

3: algorithm flow of the thesis

Saliency region selection in large aerial imagery using multi-scale SLIC Segmentation