• Laser & Optoelectronics Progress
  • Vol. 58, Issue 22, 2211003 (2021)
Haoyue Liu*, Wenwei Ma, Xiao Fu, Chengxiu Shen, and Yaling Wang
Author Affiliations
  • Internet Finance Laboratory, TK.CN Insurance Co., Ltd., Wuhan, Hubei 430014, China
  • show less
    DOI: 10.3788/LOP202158.2211003 Cite this Article Set citation alerts
    Haoyue Liu, Wenwei Ma, Xiao Fu, Chengxiu Shen, Yaling Wang. Image Manipulation Detection Algorithm Based on Improved RGB-N[J]. Laser & Optoelectronics Progress, 2021, 58(22): 2211003 Copy Citation Text show less
    Structure of improved RGB-N model
    Fig. 1. Structure of improved RGB-N model
    SRM filter. (a) KB kernel; (b) KV kernel; (c) second order linear kernel
    Fig. 2. SRM filter. (a) KB kernel; (b) KV kernel; (c) second order linear kernel
    Structure of feature extraction network
    Fig. 3. Structure of feature extraction network
    Self-attention module
    Fig. 4. Self-attention module
    Authenticity judgement module
    Fig. 5. Authenticity judgement module
    Schematic diagram of positive and negative sample selection method. (a) Spliced image; (b) source image of mosaic target; (c) manipulation target label
    Fig. 6. Schematic diagram of positive and negative sample selection method. (a) Spliced image; (b) source image of mosaic target; (c) manipulation target label
    Training samples and labeled samples. (a) Target splicing; (b) target erasure and repair
    Fig. 7. Training samples and labeled samples. (a) Target splicing; (b) target erasure and repair
    Visual effects of training different filter parameters. (a) None; (b) KB kernel; (c) second order linear kernel; (d) KB kernel and second order linear kernel
    Fig. 8. Visual effects of training different filter parameters. (a) None; (b) KB kernel; (c) second order linear kernel; (d) KB kernel and second order linear kernel
    Visual outputs of model. (a) Target splicing; (b) target erasure and repair; (c) normal images
    Fig. 9. Visual outputs of model. (a) Target splicing; (b) target erasure and repair; (c) normal images
    AlgorithmF1 scoreM /%
    ELA0.2351.8
    DCT0.4342.3
    NOI10.2871.9
    Faster-RCNN0.57016.1
    EXIF-Consistency0.6833.7
    RGB-N0.72216.7
    Proposed algorithm0.7590.2
    Table 1. Results of each algorithm
    AlgorithmQF 100QF 90QF 80QF 70QF 60QF 50
    ELA0.2350.2310.2290.2230.2150.207
    DCT0.4340.2050.1980.1850.1030.096
    NOI10.2870.2850.2810.2740.2580.235
    Faster-RCNN0.5700.5700.5670.5640.5590.550
    EXIF-Consistency0.6830.6780.6770.6710.6610.653
    RGB-N0.7220.7220.7190.7160.7130.708
    Proposed algorithm0.7590.7590.7590.7540.7420.738
    Table 2. F1 score of each algorithm when using different quality factors for compression
    Training kernelF1 scoreM /%
    None0.71817.9
    KB kernel0.72216.7
    Second order kernel0.68418.5
    KB and second order kernel0.65919.3
    Table 3. Results of training different filter parameters
    Self-attention moduleAverage poolingAuthenticity judgement moduleSample selectionF1 scoreM /%
    LossLoss and output
    Faster-RCNN0.57016.1
    0.57211.6
    Proposed algorithm0.72216.7
    0.75316.9
    0.75416.0
    0.75911.5
    0.7583.5
    0.7590.2
    Table 4. Results of ablation experiment
    Average pooling layerF1 scoreM /%
    None0.75316.9
    U10.75416.1
    U1+U20.75416.0
    U1+U2+U30.74617.1
    U1+U2+U3+U40.74117.8
    Table 5. Results of noise network pooling method
    Loss functionF1 scoreM /%
    BCE0.7530.8
    Dice loss0.7570.2
    BCE+Dice Loss0.7590.2
    Table 6. Results of different loss functions
    Haoyue Liu, Wenwei Ma, Xiao Fu, Chengxiu Shen, Yaling Wang. Image Manipulation Detection Algorithm Based on Improved RGB-N[J]. Laser & Optoelectronics Progress, 2021, 58(22): 2211003
    Download Citation