2017-06-03 Hierarchical Image Saliency Detection on Extended CSSD

Lin Yao

The learning of " Hierarchical Image Saliency Detection on Extended CSSD" (J. Shi, Q. Yan, Li Xu, and J. Jia, IEEE Transactions on Pattern Analysis and Machine Intelligence, Apr. 2016, pp.717-729)

Background

When objects contain salient small-scale patterns, saliency could generally be misled by their complexity. Aiming to solve this notorious and universal problem, this paper proposes a hierarchical framework, to analyze saliency cues from multiple levels of structure, and then integrate them for the final saliency map through hierarchical inference.

Model

1. Layer Generation

1. To produce the three layers, we first generate an initial over segmentation by the watershed-like method.

Use the Gradient Magnitude as the Segmentation Function, The gradient is high at the borders of the objects and low (mostly) inside the objects. Segment the image to get initial map by using the watershed transform.

Grad=Ix2+Iy2

Where Ix indexes the gradient in the horizontal direction, and Iy indexes the gradient in the vertical direction.

1. 
   In fact, a large pixel number does not necessarily correspond to a large-scale region in human perception. This model define a new encompassment scale measure based on shape uniformities and use it to obtain region sizes in the merging process.

scaleR=argmintRt×t|Rt×t⊆R

1. 
   Compute Region Scale

minyDty|y∈Ri>0

Dt=M-kt⋆M , where kt indexes a mean filter of size t×t, and y indexes pixels. It is based on the observ