2017-05-16 Saliency Detection with Spaces of Background-Based Distribution

The learning of "Saliency Detection with Spaces of Background-Based Distribution" (T. Zhao, L. Li, X.-H. Ding. IEEE Signal Processing Letters, May, 2016, pp. 683-687)

Create four spaces of background-based distribution to model the background in PCA space, then calculate mahalanobis distances between all the patches and background to get saliency map.

1.  Model overview
   1. Segment the image into 7*7 patches, extract features: RGB, I-RG-BY, then every patch will be represented by a 294*1 column vector (6*7*7=294).
   2. Get four groups of patches that are located around the image border denoted as Gqq=14 , where G1=BtBl  , G2=BtBr  , G3=BbBl  , G4=BbBr
   3. Create four spaces of background-based distribution to model the background in PCA space.

Unlike DSR, the distance between each patch and the background is computed directly in PCA space instead of computing the reconstruction error of each segment.

1. Construct covariance matrices C of each background group.
2. Eigendecomposition get the matrices of eigenvectors Uq  and eigenvaluesΛq .
3. The feature vector of each block is realized to projection transformation

tqi=UqTfi

1. Compute a distance dq(i)  as its Mahalanobis distance via

dqi=Λq-12tqi-pq0

pq0=x' , average features of every background group. Weighted distance calculation is performed in each direction. A patch which is similar to the background should be highly probable in the distribution while a patch belongs to the salient region should be highly probable away from the distribution.

 5）dqi<H , dqi=0,  where H is set to be the average value of dq  .Get Sqthq=14

1. 1. Create the single-scale coarse saliency map by taking a weighted average of the above four maps.
2. Calculate the entropy values of Sqthq=14 to be etrqq=14 , respectively.

etrq=Hgn⋆Sqth=-i=1npilogpi

gn : Gaussian filtering.

The entropy is used in “Visual saliency based on scale-space analysis in the frequency domain”. Considering the spatial geometric information, employing a Gaussian kernel to filter the map, and then compute the conventional entropy on the smoothed map.

1. Get coarse saliency map.

Scs=q=14wq×Sqth

                etrq≤etrave , wq=1, otherxise,  wq=0 , etrave  is the average entropy.

1. 1. Extract the foreground region and background by directly thresholding the coarse saliency map with its mean value as the prior probability map, and calculate the posterior probability map.
   2. Refinement with Geodesic Distance

Sq=j=1Nwqj×Sj

wqj  is decided by geodesic distance between any two superpixels.

These two papers have many similarities, and in every step have done is worth reference.