Objective Assessment of Stereoscopic Image Comfort Based on Convolutional Neural Network

Sumei Li; Yongli Chang; Zhicheng Duan

doi:10.3788/AOS201838.0610003

Journals >Acta Optica Sinica >Volume 38 >Issue 6 >Page 0610003 > Article

Acta Optica Sinica
Vol. 38, Issue 6, 0610003 (2018)

Objective Assessment of Stereoscopic Image Comfort Based on Convolutional Neural Network

Sumei Li, Yongli Chang^*, and Zhicheng Duan

Author Affiliations

School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China

show less

DOI: 10.3788/AOS201838.0610003 Cite this Article Set citation alerts

Sumei Li, Yongli Chang, Zhicheng Duan. Objective Assessment of Stereoscopic Image Comfort Based on Convolutional Neural Network[J]. Acta Optica Sinica, 2018, 38(6): 0610003 Copy Citation Text

show less

Fig. 1. Architecture of CNN

Download full size

Fig. 2. Architecture of three-channel CNN

Download full size

Fig. 3. (a)-(d) Source images and (e)-(h) distorted images

Download full size

Grade	Criteria for judging image damage	Degree of comfort
5	Almost no distortion	Excellent
4	Slightly distorted but not repugnant	Good
3	General distortion and a little repugnant	Fair
2	Obviously distorted but not disgusting	Poor
1	Serious distorted and disgusting	Bad

Table 1. Suggestions on quality division of stereoscopic images

Patch_256_net	Patch_32_net	Pca_net
Conv-20(5×5)	Conv-20(5×5)	Conv-50(3×3)
Max(3×3)	Max(2×2)	Conv-50(3×3)
Conv-100(5×5)	Conv-50(5×5)	Conv-50(3×3)
Max(4×4)	Max(2×2)	Max(3×3)
Conv-100(5×5)	-	-
Conv-100(4×4)
Max(3×3)
Conv-100(3×3)
Max(3×3)
FC-2500
FC-600
FC-5

Table 2. Parameter setting of network model

Structure of CNN model	Recognition rate/%
Patch_32_net	51.67
Patch_256_net	55.00
Patch_32_256_net	66.25
PCA_net	76.75
PCA_32_net	84.50
PCA_256_net	88.25
PCA_32_256_net	94.52

Table 3. Recognition rates of test samples with different structures of CNN model

Optimization method		Recognition rate /%
	Dropout layer	LRN layer
No	Yes	93.20
Yes	No	93.40
Yes	Yes	94.52

Table 4. Influence of Dropout and LRN layer on PCA_32_256_net model recognition rate

Algorithm	Recognition rate /%	Train time /s	Test time /s
Proposed	94.52	7920.0000	0.1470
SVM^[6]	92.50	20.2700	0.0047
ELM^[7]	93.85	0.0025	0.0037

Table 5. Recognition rates of test algorithms

Sumei Li, Yongli Chang, Zhicheng Duan. Objective Assessment of Stereoscopic Image Comfort Based on Convolutional Neural Network[J]. Acta Optica Sinica, 2018, 38(6): 0610003

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information