• Opto-Electronic Engineering
  • Vol. 51, Issue 9, 240139-1 (2024)
Bin Wang1, Yongqiang Bai2, Zhongjie Zhu2, Mei Yu1,*, and Gangyi Jiang1
Author Affiliations
  • 1Faculty of Information Science and Engineering, Ningbo University, Ningbo, Zhejiang 315211, China
  • 2College of Information and Intelligent Engineering, Zhejiang Wanli University, Ningbo, Zhejiang 315100, China
  • show less
    DOI: 10.12086/oee.2024.240139 Cite this Article
    Bin Wang, Yongqiang Bai, Zhongjie Zhu, Mei Yu, Gangyi Jiang. No-reference light field image quality assessment based on joint spatial-angular information[J]. Opto-Electronic Engineering, 2024, 51(9): 240139-1 Copy Citation Text show less
    Different representations of light field image. (a) MLI; (b) SAIs
    Fig. 1. Different representations of light field image. (a) MLI; (b) SAIs
    Overall framework of SAE-BLFI
    Fig. 2. Overall framework of SAE-BLFI
    Schematic diagram of spatial-angular separation
    Fig. 3. Schematic diagram of spatial-angular separation
    EPI under different distortion conditions for two scenarios
    Fig. 4. EPI under different distortion conditions for two scenarios
    Boxplot of SROCC distribution in K-fold cross-validation on Win5-LID and NBU-LF1.0 datasets. (a) Win5-LID; (b) NBU-LF1.0
    Fig. 5. Boxplot of SROCC distribution in K-fold cross-validation on Win5-LID and NBU-LF1.0 datasets. (a) Win5-LID; (b) NBU-LF1.0
    F-test statistical significance analysis on Win5-LID and NBU-LF1.0 datasets. (a) Win5-LID; (b) NBU-LF1.0
    Fig. 6. F-test statistical significance analysis on Win5-LID and NBU-LF1.0 datasets. (a) Win5-LID; (b) NBU-LF1.0
    TypesMethodsWin5-LIDNBU-LF1.0SHU
    PLCC↑SROCC↑RMSE↓PLCC↑SROCC↑RMSE↓PLCC↑SROCC↑RMSE↓
    NR 2D-IQABRISQUE[24]0.62170.45370.76040.49890.38710.78790.90110.88830.4591
    GLBP[25]0.53570.41500.81300.50560.34900.76470.71680.65650.7504
    FR LF-IQAMDFM[5]0.77630.74710.62490.78880.75590.56490.89470.89080.4863
    Min's[7]0.72810.66450.68740.71040.65790.64390.84970.84700.5757
    Meng's[26]0.69830.63470.72030.84040.78250.48890.92790.92030.4039
    NR LF-IQABELIF[27]0.57510.50590.78650.70140.63890.62760.89670.86560.4803
    NR-LFQA[8]0.72980.69790.62710.85280.81130.46580.92240.92290.4132
    Tensor-NLFQ[12]0.58130.48850.77060.68840.62460.63050.93070.90610.3857
    VBLFI[10]0.72130.67040.68430.80270.75390.52180.92350.89960.4064
    4D-DCT-LFIQA[2]0.82340.80740.54460.83950.82170.48710.94000.93200.3691
    DeeBLiF[18]0.84270.81860.51600.85830.82290.45880.95480.94190.3185
    SATV-BLiF[28]0.79330.77040.58420.85150.82370.46860.93320.92840.3897
    Proposed0.86530.84510.48630.91080.89370.36580.96490.95470.2808
    Table 1. Overall performance comparison of different methods on different LFI datasets
    TypesMethodsWin5-LIDNBU-LF1.0Hit count
    HEVCJEPG2KLNNNNNBIEPICNNMDRVDSR
    NR 2DIQABRISQUE[24]0.56410.78010.52220.24620.34350.41450.57950.43310.79370
    GLBP[25]0.71650.48530.46780.30110.32290.39950.43440.44780.73810
    FR LF-IQAMDFM[5]0.79220.76690.64370.66920.80250.90890.78990.73860.87091
    Min's[7]0.69970.65070.61590.62880.81560.86670.73610.79630.93761
    Meng's[26]0.88860.69390.84590.80010.74290.90180.79970.57830.92252
    NR LF-IQABELIF[27]0.76660.63790.60970.54520.76800.71220.68740.61280.79890
    NR-LFQA[8]0.75710.73380.63620.70260.89300.88070.76530.61110.81640
    Tensor-NLFQ[12]0.68530.57990.56630.58970.69460.72030.52450.54170.80180
    VBLFI[10]0.71410.74490.69080.71970.83160.83720.71950.46130.91340
    4D-DCT-LFIQA[2]0.86980.89460.81270.82350.90400.87190.71000.80950.88822
    DeeBLiF[18]0.96480.81950.79280.83060.91840.88760.72480.69610.88573
    SATV-BLiF[28]0.79180.86850.75660.85250.92820.91900.77220.64980.86172
    Proposed0.94170.89550.84720.87420.91650.91530.77490.84430.92947
    Table 2. SROCC values for different distortion types across various methods on Win5-LID and NBU-LF1.0 datasets
    Win5-LIDNBU-LF1.0
    PLCCSROCCRMSEPLCCSROCCRMSE
    SF0.83380.81790.52170.88530.86540.4104
    AF0.81950.79920.53220.87090.85660.4208
    EF0.79500.79920.56520.86370.84780.4176
    SF+AF0.85130.82850.50570.90570.88570.3845
    SF+AF+EF0.86530.84510.48630.91080.89370.3658
    Table 3. Ablation experiments of different functional modules on Win5-LID and NBU-LF1.0 datasets
    MethodsPlatformDeviceTime/s
    BELIF[27]MatlabCPU167.60
    NR-LFQA[8]MatlabCPU220.92
    Tensor-NLFQ[12]MatlabCPU630.47
    VBLFI[10]MatlabCPU68.48
    4D-DCT-LFIQA[2]MatlabCPU148.29
    DeeBLiF[18]PytorchGPU2.77
    SATV-BLiF[28]MatlabCPU4.38
    ProposedPytorchGPU3.77
    Table 4. Comparison of running time for different NR LF-IQA methods
    MethodsNBU-LF1.0 (NN)SHU (JPEG2000)
    PLCCSROCCPLCCSROCC
    4D-DCT-LFIQA0.77530.70400.78240.7967
    DeeBLiF0.82530.72650.76090.7252
    Proposed0.90820.86100.88210.8717
    Table 5. The results of training the model on the Win5-LID dataset and testing it on the NBU-LF1.0 and SHU datasets
    Bin Wang, Yongqiang Bai, Zhongjie Zhu, Mei Yu, Gangyi Jiang. No-reference light field image quality assessment based on joint spatial-angular information[J]. Opto-Electronic Engineering, 2024, 51(9): 240139-1
    Download Citation