《Saliency Filters: Contrast Based Filtering for Salient Region Detection》阅读理解

作者:Federico Perazzi

作者:Philipp Kr 等 论文地址: IEEE Xplore Full-Text PDF: 发表于2012 CVPR   论文主要包括四个部分: Abstraction、Element uniqueness、Element distribution、Saliency assignment。   第一,Abstraction。 aim to decompose the image into basic elements that preserve relevant structure, but abstract undesirable detail. 具体操作即edge-preserving, localized oversegmentation based on color,其用SLIC超像素分割来实现,效果如图2所示。

《Saliency Filters: Contrast Based Filtering for Salient Region Detection》阅读理解

第二, Element uniqueness——利用颜色信息。 即元素的“rarity”,其定义为:

《Saliency Filters: Contrast Based Filtering for Salient Region Detection》阅读理解

能够有效结合全局和局部的对比度,局部对比度强调边界,w约等于1时是全局的,对局部地必读变化不敏感。这里w是高斯加权,能够将原有《Saliency Filters: Contrast Based Filtering for Salient Region Detection》阅读理解运算降为线性的《Saliency Filters: Contrast Based Filtering for Salient Region Detection》阅读理解《Saliency Filters: Contrast Based Filtering for Salient Region Detection》阅读理解《Saliency Filters: Contrast Based Filtering for Salient Region Detection》阅读理解

式(1)可写为:

《Saliency Filters: Contrast Based Filtering for Salient Region Detection》阅读理解

第三,Element distribution——利用空间分布。

图像背景信息分布在整个图像,因此方差较大,而显著性信息则比较聚集。

《Saliency Filters: Contrast Based Filtering for Salient Region Detection》阅读理解

这里,同样定义了一个高斯weight,《Saliency Filters: Contrast Based Filtering for Salient Region Detection》阅读理解。式(3)

可改写为:

《Saliency Filters: Contrast Based Filtering for Salient Region Detection》阅读理解

第四,Saliency assignment。

 

We start by normalizing both uniqueness U i and distribution D i to the range [0 .. 1] .

《Saliency Filters: Contrast Based Filtering for Salient Region Detection》阅读理解

论文指出,Di的贡献更大: In practice we found the distribution measure D i to be of higher signifificance and discriminative power. Therefore, we use an exponential function in order to emphasize D i . 直接对每一个元素进行上采样来获得最终显著性值不能避免图像Ab straction所产生的分割错误,因此论文提出通过融合一个pixel周围的pixel来获得最终显著性值:

《Saliency Filters: Contrast Based Filtering for Salient Region Detection》阅读理解

其中,w仍然是高斯加权:《Saliency Filters: Contrast Based Filtering for Salient Region Detection》阅读理解《Saliency Filters: Contrast Based Filtering for Salient Region Detection》阅读理解,能够保证对 local and color sensitive.

In a final step we rescale the saliency map to the range [0 .. 1] or to contain at least 10% saliency pixels.   实验结果:

《Saliency Filters: Contrast Based Filtering for Salient Region Detection》阅读理解

《Saliency Filters: Contrast Based Filtering for Salient Region Detection》阅读理解

 
上一篇:Python图像处理(三)--从图像拼接到视频拼接与整合处理


下一篇:《Group Collaborative Learning for Co-Salient Object Detection》——CVPR2021 Co-salient object detection