Saliency Region Selection in Large Aerial Imagery using Multi-scale SLIC Segmentation

最新推荐文章于 2020-01-23 22:10:50 发布

原创最新推荐文章于 2020-01-23 22:10:50 发布 · 1.1k 阅读

2 ·

CC 4.0 BY-SA版权

Old 专栏收录该内容

75 篇文章

订阅专栏

本文提出了一种针对大型航拍图像的显著区域选择方法，该方法利用了多尺度SLIC分割技术。当背景区域较大且较为均匀时，超像素会呈现平均大小和规则形状；而遇到感兴趣的对象时，超像素则紧随对象边界，表现出不规则形状。通过多尺度SLIC处理，可以有效识别不同大小的目标。

1 : Saliency Region Selection in Large Aerial Imagery using Multi-scale SLIC Segmentation, Proc. SPIE 8360, Airborne Intelligence, Surveillance, Reconnaissance (ISR) Systems and Applications IX, 2012

2 :

此论文中值得留意的几个观点：

(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 various sizes and strongly 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 various 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 by the Hausdorff distance defined in the superpixel string-to-string presentation.

3: 论文算法流程