Salient Detection Based on Cascaded Convolutional Neural Network

Fig. 4. Comparison of saliency maps. (a) Original images; (b) GT; (c) proposed algorithm; (d) SRM algorithm; (e) Amulet algorithm; (f) UCF algorithm; (g) DCL algorithm; (h) DHS algorithm; (i) RFCN algorithm; (j) MDF algorithm; (k) BL algorithm; (l) wCO algorithm

Download full size

Fig. 5. P-R curves of different methods in different datasets. (a) ECSSD dataset; (b) SED2 dataset

Download full size

Table 1. Network parameters of BRS
View table
Table 1. Network parameters of BRS
Layer Channel Kernel size Dilation rate
1 64 (1+3)×64×3×3 1
2 64 64×64×3×3 1
3 64 64×64×3×3 1
4 128 64×128×3×3 1
5 128 128×128×3×3 2
6 128 128×128×3×3 3
7 1 128×1×3×3 1

Table 2. Size comparison of two performance indicators

View table

Table 2. Size comparison of two performance indicators

Algorithm	ECSSD		SED2
Algorithm	MAE	F-measure	MAE	F-measure
wCO	0.171	0.651	0.131	0.694
BL	0.216	0.684	0.189	0.705
MDF	0.105	0.833	0.115	0.801
RFCN	0.107	0.834	0.113	0.767
DHS	0.060	0.871	0.079	0.822
DCL	0.074	0.827	0.101	0.795
UCF	0.078	0.841	0.086	0.810
Amulet	0.059	0.869	0.080	0.871
SRM	0.056	0.892	0.076	0.894
Base-net	0.073	0.817	0.082	0.787
Base-net+PPM	0.064	0.842	0.079	0.834
Base-net+BRS	0.057	0.869	0.075	0.861
Proposed	0.052	0.898	0.070	0.902