CVPR 2011 Global contrast based salient region detection

Two salient region detection methods are proposed in this paper: HC AND RC

HC: Histogram based contrast

1. Primary method

It is simply to calculate the saliency of each color in the input image, where each pixel's saliency is defined using its color contrast to all other pixels in the image in L*a*b space:

The above equation can be expanded as,

where N is the number of pixels. Therefore we can conclude that the same color have the same saliency, so saliency value for each color is,

n is the number of distinct pixel colors,  is the probability of the corresponding pixel color in the image I.

 

2. Speeding up strategy

To reduce the number of colors from 256^3 to 12^3 = 1728, and finally to n = 85 colors in this post. 

3. Color space smoothing

In order to reduce noisy saliency results caused by such randomness. We replace the saliency value of each color by the weighted average of the saliency values of similar colors. We choose m = n/4 nearsest colors to refine each color.

See the original paper for the detail of this equation.

 

 

RC: Region based contrast

1. Segment the input image into regions using [45](see the original paper)

2. Build the color histogram for each region

3. For a region r_k, we compute its saliency value by measuring its color contrast to all other regions in the image,

where  is the number of pixels of region r_i, designed to emphasize color contrast to bigger regions, and 

f(c_k,i) is the probability of the i-th color among all n_k colors in the k-th region, used to emphasize the color differences between dominant colors.

4. Spatially weighted region contrast

We now can incorporate spatial information to the above equation to increase the effects of closer regions and decrease the farther ones.

 is the spatial distance between the two regions. controls the strength of spatial information, the bigger value make less effect of the close regions relatively,  is a spatial prior weighting term similar to center bias.

 

 

5. Iteratively segmentation using graph-cut.

Dilation and erosion after each iteration. The region inside the eroded region is set to foreground, and the remaining areas are set to unknown.

 

 However, the (f) is not understood...  ≡(▔﹏▔)≡

转载于:https://www.cnblogs.com/guanyu-zuike/p/6223507.html