Paper reading :Hierarchical Saliency Detection

Overview of hierachical framework-2013

---

Image Layer Extraction

—————coarse representation of the input with different degrees of details

1. define scale of region R

2. Efficient Algorithm to Compute Region Scale

避免穷举法计算带来的时间代价，采用box filter with t x t kernel to the segment and cheking if all pixel values inside the segment are changed during filtering.

3. Layer Generation

First, over segmentation : by the watershed like method. For each segmented region, we compute a scale value

Single-Layer Saliency Cues

———————saliency cues are applied to find important pixels from the perspectives of color, position and size

1. Local contrast

c_i and c_j are colors of regions R_i and R_j respectively. ( Lab color space)

w(R_j) : the number of pixels in R_j.

φ(i, j): is set to exp{−D(R_i, R_j)/σ²}, Parameter σ² is set to the product of (0.2)² and the particular scale threshold for the current layer.

2. Location heuristic——pixels close to a natural image center could be salient in many cases

where {x0, x1 · · · } is the set of pixel coordinates in region R_i;

λ is set to 9 in general (控制);

x_c is the coordinate of the image center

local contrast and location cues normalized to range [0,1)

3. Initial saliency value

initial saliency cue maps in three layers

---

Hierarchical Inference——in 2013

• 多层融合时不采用简单的加权，而是用分层推导的方法得出能量方程，最后采用置信传播算法最小化能量方程。

• For a node corresponding to region i in layer L_l, we define a saliency variable s^l_i.

• minimize the following energy function:

• Data term E_D(s^l_i) is to gather separate saliency confidence.

  

  • β^l controls the layer confidence

  

• Hierarchy term E_S(s^l_i,s^l+1_j) enforce consistency between corresponding regions in different layers.

  (notice: R^l_i and R^l+1_j are corresponding in two layers, we have R^l_i ⊆ R^l+1_j)

  

  Finally, minimize Eq.(5) by belief propagation.

Overview of hierachical framework-2013

• 对比 :

• 13年:对于L层i节点显著性值考虑与初始显著性值差距最小化，同时与L+1层对应父节点显著性值差异最小化；

• 15年:在13年的基础上，加上一条规则：对于L层i节点考虑与L层相邻的所有节点 j 依据相似性缩小差距（相似度大，差距越小）,用来平滑相似的相邻区间的显著性值

• 效果上的前后对比：
  

---

Hierarchical Inference——in 2015

• new energy function :

• Data term E_D(s^l_i) and Hierarchy term E_H(s^l_i,s^l+1_j) 几乎与2013年的一样，不同之处在于添加了权重参数：

βⁱ is fixed as {0.5, 4, 2} for i = {1, 2, 3};

• Local Consistency term E_C(s^l_i,s^l+1_j) which enforces intra-layer smoothness. It is used to make saliency assignment smooth between adjacent similar regions.

Notation A(R^k_i):R^k_i的邻域集合。which means R^k_j∈ A(R^k_i) if R^k_j is the adjacent region of R^k_i

γ^k determines the strength of consistency for each layer.

(in the LUV color space)

σ_c is the normalization parameter.

Finally, optimaization.(有点儿复杂，以后再写，而且文中有很多的参数还没太弄明白，以后再更新)