Deep Network Saliency Detection Based on Global Model and Local Optimization

Feng Liu; Tongsheng Shen; Shuli Lou; Bing Han

doi:10.3788/AOS201737.1215005

Journals >Acta Optica Sinica >Volume 37 >Issue 12 >Page 1215005 > Article

Acta Optica Sinica
Vol. 37, Issue 12, 1215005 (2017)

Deep Network Saliency Detection Based on Global Model and Local Optimization

Feng Liu^1、*, Tongsheng Shen², Shuli Lou¹, and Bing Han³

Author Affiliations

¹ Department of Control Engineering, Naval Aeronautical and Astronautical University, Yantai, Shandong 264001, China

² China Defense Science and Technology Information Center, Beijing 100142, China

³ Element 98 of Unit 92493, PLA, Huludao, Liaoning 125000, China

show less

DOI: 10.3788/AOS201737.1215005 Cite this Article Set citation alerts

Feng Liu, Tongsheng Shen, Shuli Lou, Bing Han. Deep Network Saliency Detection Based on Global Model and Local Optimization[J]. Acta Optica Sinica, 2017, 37(12): 1215005 Copy Citation Text

show less

Fig. 1. Structure of model

Download full size

Fig. 2. Network structure of global model

Download full size

Fig. 3. Generation of feature vectors of super-pixel

Download full size

Fig. 4. Saliency maps of different steps. (a) Original images; (b) global models; (c) images of local optimization; (d) final saliency maps; (e) ground truth images

Download full size

Fig. 5. PR curves of four data sets with different methods. (a) SOD; (b) PASCAL-S; (c) CSSD; (d) MSRA

Download full size

Fig. 6. Visual comparisons of our results and others. (a) Original images; (b) ground truth images; (c) proposed method; (d) LEGS; (e) DRFI; (f) HDCT; (g) wCtr; (h) PCA; (i) GBVS

Download full size

Color texture feature			Differential feature
Feature	Descriptor	Dimension	Definition	Dimension
Average RGB value	a₁	3	d( $\begin{matrix} {a^{R}}_{1} \end{matrix}$ , $\begin{matrix} a_{1}^{I} \end{matrix}$ )	3
Average lab value	a₂	3	d( $\begin{matrix} {a^{R}}_{2} \end{matrix}$ , $\begin{matrix} {a^{I}}_{2} \end{matrix}$ )	3
Gabor filter response	r	24	d(r^R,r^I)	24
Maximum Gabor response	r	1	d(r^R,r^I)	1
RGB color histogram	h₁	24	χ²( $\begin{matrix} {h^{R}}_{1} \end{matrix}$ , $\begin{matrix} {h^{I}}_{1} \end{matrix}$ )	1
Lab color histogram	h₂	24	χ²( $\begin{matrix} {h^{R}}_{2} \end{matrix}$ , $\begin{matrix} {h^{I}}_{2} \end{matrix}$ )	1
HSV color histogram	h₃	24	χ²( $\begin{matrix} {h^{R}}_{3} \end{matrix}$ , $\begin{matrix} {h^{I}}_{3} \end{matrix}$ )	1

Table 1. Feature vectors of contrast descriptor

Feature	Dimension	Feature	Dimension
Normalized x of regional center	1	Regional connectivity^[16]	1
Normalized y of regional center	1	RGB color variance	3
Normalized are	1	Lab color variance	3
Aspect ratio of bounding box	1	HSV color variance	3
Bounding box width	1	Gabor filter response variance	24
Bounding box length	1	Normalized area of neighborhood	1

Table 2. Parameters of regional feature descriptor

Dataset	Curve	Proposed method	LEGS method	DRFI method	HDCT method	wCtr method	GBVS method	PCA method	GC method	SF method	MR method
SOD	F-measure	73.1	67.4	70.2	65.4	63.7	61.3	54.9	50.6	55.3	54.2
SOD	MAE	20.4	21.2	24.1	26.6	24.5	26.9	25.3	28.8	26.7	27.4
CSSD	F-measure	84.6	83.1	78.8	70.5	66.8	65.3	57.5	55.7	54.5	67.5
CSSD	MAE	12.8	11.9	17.9	19.9	18.4	22.7	25.2	23.4	20.1	19.0
PASCAL	F-measure	75.3	74.9	69.9	60.4	61.1	69.3	53.1	61.6	57.4	58.3
PASCAL	MAE	14.7	15.5	20.3	22.9	20.1	17.8	23.9	25.5	21.4	21.2
MSRA	F-measure	91.2	90.5	91.9	80.5	78.3	65.9	70.1	68.2	62.5	78.3
MSRA	MAE	10.4	8.9	14.3	11.9	16.6	13.7	18.9	14.7	16.2	13.0

Table 3. Comparison of F-measure scores with different methods%

Feng Liu, Tongsheng Shen, Shuli Lou, Bing Han. Deep Network Saliency Detection Based on Global Model and Local Optimization[J]. Acta Optica Sinica, 2017, 37(12): 1215005

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information