• Laser & Optoelectronics Progress
  • Vol. 55, Issue 12, 121004 (2018)
Dongzhen Huang1、2, Qin Zhao1、2, Huawei Liu1, Baoqing Li1, and Xiaobing Yuan1、*
Author Affiliations
  • 1 Key Laboratory of Microsystem Technology, Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, Shanghai 201800, China
  • 2 University of Chinese Academy of Sciences, Beijing 100049, China
  • show less
    DOI: 10.3788/LOP55.121004 Cite this Article Set citation alerts
    Dongzhen Huang, Qin Zhao, Huawei Liu, Baoqing Li, Xiaobing Yuan. Dense Disparity Map Extraction Method Based on Improved Convolutional Neural Network[J]. Laser & Optoelectronics Progress, 2018, 55(12): 121004 Copy Citation Text show less
    Dual pyramid network structure
    Fig. 1. Dual pyramid network structure
    Comparison of network structures
    Fig. 2. Comparison of network structures
    Relationship between error rate and parameters of additional convolutional layers. (a) Error rate with different convolutional kernel sizes; (b) test error with different number of convolutional layers
    Fig. 3. Relationship between error rate and parameters of additional convolutional layers. (a) Error rate with different convolutional kernel sizes; (b) test error with different number of convolutional layers
    Comparison of disparity maps. (a) Signpost 1; (b) fence; (c) car; (d) signpost 2; (e) traffic sign; (f) signpost 3
    Fig. 4. Comparison of disparity maps. (a) Signpost 1; (b) fence; (c) car; (d) signpost 2; (e) traffic sign; (f) signpost 3
    Zoomed parts over region marked by box in Fig. 4. (a) Signpost 1; (b) fence; (c) car; (d) signpost 2; (e) traffic sign; (f) signpost 3
    Fig. 5. Zoomed parts over region marked by box in Fig. 4. (a) Signpost 1; (b) fence; (c) car; (d) signpost 2; (e) traffic sign; (f) signpost 3
    Effect of structure on subjective quality. (a) Left input image; (b) without additional convolutional layers; (c) without dual pyramid structure; (d) default
    Fig. 6. Effect of structure on subjective quality. (a) Left input image; (b) without additional convolutional layers; (c) without dual pyramid structure; (d) default
    ItemMC-CNN-fstProposed network
    3×35×5
    Time consumption /s0.19020.25610.2438
    Test error /%3.0292.7842.795
    Table 1. Comparison of complexity and accuracy
    ImageMC-CNN-fstProposed (5×5)
    10.8070.807
    24.1803.892
    31.2101.147
    41.2071.117
    58.9608.006
    66.2285.866
    71.5881.475
    80.2930.213
    92.9342.658
    103.0202.877
    Average3.0292.795
    Table 2. Objective index of 10 test images
    ItemMC-CNN-fstProposed(5×5)
    Complexity ofconvolution layers111168221760
    Time consumption /s0.19020.2438
    Whole timeconsumption /s0.40400.4770
    Table 3. Comparison of complexity
    ItemMC-CNN-fstMC-CNN-fst(1×1)Without additionalconvolutional layersWithout dualpyramid structureDefault
    3 pixel /%3.0293.0173.0452.7572.795
    Time consumption /s0.19020.20810.22510.22520.2438
    Table 4. Comparison of different network structures
    Dongzhen Huang, Qin Zhao, Huawei Liu, Baoqing Li, Xiaobing Yuan. Dense Disparity Map Extraction Method Based on Improved Convolutional Neural Network[J]. Laser & Optoelectronics Progress, 2018, 55(12): 121004
    Download Citation