Depth Estimation for Phase-Coding Light Field Based on Neural Network

Chengzhuo Yang; Sen Xiang; Huiping Deng; Jing Wu

doi:10.3788/LOP221145

Journals >Laser & Optoelectronics Progress >Volume 60 >Issue 12 >Page 1211002 > Article

Laser & Optoelectronics Progress
Vol. 60, Issue 12, 1211002 (2023)

Depth Estimation for Phase-Coding Light Field Based on Neural Network

Chengzhuo Yang^1,2, Sen Xiang^1,2,*, Huiping Deng^1,2, and Jing Wu^1,2

Author Affiliations

¹School of Information Science and Engineering, Wuhan University of Science and Technology, Wuhan 430081, Hubei, China

²Engineering Research Center for Metallurgical Automation and Measurement Technology, Ministry of Education, Wuhan University of Science and Technology, Wuhan 430081, Hubei, China

show less

DOI: 10.3788/LOP221145 Cite this Article Set citation alerts

Chengzhuo Yang, Sen Xiang, Huiping Deng, Jing Wu. Depth Estimation for Phase-Coding Light Field Based on Neural Network[J]. Laser & Optoelectronics Progress, 2023, 60(12): 1211002 Copy Citation Text

show less

Fig. 1. Three-branch light field depth estimation network

Download full size

Fig. 2. Multi-scale feature extraction module

Download full size

Fig. 3. Data augmentation by rotation

Download full size

Fig. 4. Convergence curve of network training

Download full size

Fig. 5. Comparison of the data results of each method on the test set. (a) MAE; (b) BP7; (c) BP5; (d) BP3

Download full size

Fig. 6. Comparison of depth map results of various methods. (a) CAE; (b) OCC; (c) SPO; (d)REFOCUS; (e) EPINet; (f) proposed method; (g) ground-truth

Download full size

Fig. 7. Comparison of error maps of the proposed method and EPINet

Download full size

Fig. 8. Comparison of depth map results of the proposed network with/without center view

Download full size

Fig. 9. Comparison of error map results of the proposed network with/without center view

Download full size

Method	Scene 1									Scene 8						Scene 24
Method	MAE /10²		BP7 /%		BP5 /%			BP3 /%		MAE /10²		BP7 /%	BP5 /%		BP3 /%	MAE /10²		BP7 /%	BP5 /%		BP3 /%
CAE	34.86		1.371		1.666			1.959		29.09		0.668	0.864		1.122	78.29		4.769	5.187		5.952
OCC	77.36		2.498		3.923			5.282		69.50		1.295	2.362		4.051	88.43		4.131	5.618		9.697
SPO	145.38		1.002		18.430			18.430		194.36		1.581	28.213		53.186	90.13		0.947	2.634		11.222
REFOCUS	91.42		4.710		6.350			8.489		68.36		1.761	3.289		5.437	89.05		4.980	6.424		9.599
EPINet	16.99		0.021		0.052			0.193		15.26		0.013	0.056		0.206	19.79		0.191	0.383		0.858
Proposed method	17.66		0.043		0.091			0.091		18.11		0.015	0.055		0.183	20.72		0.187	0.380		0.815
Method	Scene 38									Scene 40						Scene 55
Method	MAE /10²		BP7 /%			BP5 /%			BP3 /%	MAE /10²	BP7 /%			BP5 /%	BP3 /%	MAE /10²	BP7 /%			BP5 /%		BP3 /%
CAE	31.63		1.127		1.290			1.544		25.76		0.927	1.108		1.388	31.74		0.886	1.094		1.415
OCC	82.35		2.218		4.930			7.082		68.27		2.990	4.712		5.893	78.64		1.630	3.252		5.793
SPO	140.89		0.785		19.203			32.404		110.02		1.025	19.219		28.477	198.86		1.006	28.920		55.598
REFOCUS	81.27		3.615		5.396			6.995		97.72		7.074	8.846		10.340	75.81		2.605	4.483		7.101
EPINet	22.34		0.071		0.156			0.409		19.02		0.055	0.147		0.370	17.30		0.024	0.066		0.252
Proposed method	22.76		0.091		0.170			0.384		16.29		0.057	0.145		0.400	19.43		0.027	0.078		0.239
Method	Scene 58									Scene 69						Scene 86
Method	MAE /10²	BP7 /%		BP5 /%			BP3 /%			MAE /10²	BP7 /%			BP5 /%	BP3 /%	MAE /10²	BP7 /%			BP5 /%		BP3 /%
CAE	29.82		0.965		1.120			1.345		35.05		1.254	1.394		1.650	25.94		0.665	0.827		1.115
OCC	82.84		1.976		4.499			6.174		67.74		1.819	2.902		4.330	74.92		1.812	3.910		5.661
SPO	208.54		1.248		38.245			56.931		192.19		1.125	30.187		52.479	189.23		1.021	32.608		51.655
REFOCUS	77.55		3.202		5.779			7.803		73.69		3.289	5.175		7.853	82.28		3.397	5.688		8.005
EPINet	19.47		0.040		0.109			0.249		15.28		0.080	0.142		0.336	16.97		0.023	0.065		0.197
Proposed method	22.29		0.055		0.118			0.282		17.68		0.077	0.142		0.332	19.33		0.018	0.064		0.191

Table 1. Comparison of objective metrics of different algorithms in different scenes

Method	MAE /10²	BP7 /%	BP5 /%	BP3 /%
CAE	34.9138	1.1345	1.3402	1.6896
OCC	74.1117	1.9221	3.5428	5.4535
SPO	161.6611	0.9002	22.5860	40.8998
REFOCUS	78.2058	3.4325	5.1551	7.2922
EPINet	18.1747	0.0648	0.1396	0.3362
Proposed method	19.7986	0.0885	0.1591	0.3551

Table 2. Comparison of average objective metrics of different algorithms

Method	Average time/s	Number of parameters
CAE	806.5161
OCC	19.3286
SPO	283.0965
REFOCUS	166.4036
EPINet	0.5533	5，124，281
Proposed method	0.2282	1，402，065

Table 3. Efficiency comparison of different algorithms

Method	MAE /10²	BP7 /%	BP5 /%	BP3 /%
Proposed method（without center view）	41.4286	0.0894	0.1702	0.3657
Proposed method（with center view）	19.7986	0.0885	0.1591	0.3551

Table 4. Comparison of average objective metrics of the proposed network with/without center view

Chengzhuo Yang, Sen Xiang, Huiping Deng, Jing Wu. Depth Estimation for Phase-Coding Light Field Based on Neural Network[J]. Laser & Optoelectronics Progress, 2023, 60(12): 1211002

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information

微信扫一扫：分享

微信扫一扫：分享