Just Noticeable Difference Estimation For Images with Structural Uncertainty WU Jinjian Xidian University
Outline Concept of JND Existing Works Inspiration JND Estimation Experimental Results Conclusion
Original image JND noise (PSNR=26.4) contaminated image
Concept of Just Noticeable Difference Original image JND noise (PSNR=26.4) contaminated image +
Concept of Just Noticeable Difference Original image JND noise (PSNR=26.4) contaminated image +
Concept of Just Noticeable Difference Original image JND noise (PSNR=26.4) contaminated image Just-noticeable difference (JND): below which the change cannot be detected by the majority (e.g., 75%) of viewers; JND is useful in perceptual based signal processing, e.g., image/video coding, watermarking, quality assessment, etc.
Factors determine JND threshold : Luminance adaptation: the HVS presents different resolutions for different background illuminations ; Existing works
Factors determine JND threshold : Luminance adaptation: the HVS presents different resolutions for different background illuminations ; Existing works
Factors determine JND threshold : Luminance adaptation: the HVS presents different resolutions for different background illuminations ; Existing works
Factors determine JND threshold : Luminance adaptation: the HVS presents different resolutions for different background illuminations ; Existing works
Factors determine JND threshold : Luminance adaptation: the HVS presents different resolutions for different background illuminations ; Existing works
Factors determine JND threshold : Contrast masking: the spatial non-uniformity of the background luminance causes masking effect;
Existing works Factors determine JND threshold : Contrast masking: the spatial non-uniformity of the background luminance causes masking effect;
Existing works Factors determine JND threshold : Contrast masking: the spatial non-uniformity of the background luminance causes masking effect;
contaminated Existing works Shortcoming of the existing works : Overestimate the orderly regions (edge); Underestimate the disorderly regions (texture); original
contaminated Existing works Shortcoming of the existing works : Overestimate the orderly regions (edge); Underestimate the disorderly regions (texture); Too much noise original
contaminated Existing works Shortcoming of the existing works : Overestimate the orderly regions (edge); Underestimate the disorderly regions (texture); original
contaminated Existing works Shortcoming of the existing works : Overestimate the orderly regions (edge); Underestimate the disorderly regions (texture); Too little noise original
Luminance Adaptation Luminance Adaptation JND Threshold JND Threshold Factors determine JND threshold : Existing works Contrast Masking Contrast Masking
Luminance Adaptation Luminance Adaptation JND Threshold JND Threshold Factors determine JND threshold : Existing works Contrast Masking Contrast Masking ? ? There must be another factor, which also determines JND
Inspiration Intuitive perception : The HVS is sensitive to image regions with simple/regular structures(with small JND); Regular structures
Inspiration Intuitive perception : The HVS is sensitive to image regions with simple/regular structures(with small JND); The HVS is insensitive to image regions with complex/irregular structures (with large JND); Regular structuresirregular structures
Theoretical Support: (Free-Energy Principle) Recent perceptual science findings indicate that the HVS possesses an internal generative mechanism (IGM); Inspiration
Inspiration
Inspiration
By adjusting the internal structure of neurons By adjusting the internal structure of neurons Actively predicts the visual signal, while ignores the residual Actively predicts the visual signal, while ignores the residual When Perceiving an input scene Inspiration
Inspiration Theoretical Support: (Free-Energy Principle) According to the free-energy principle, orderly structures are easily to be predicted, and the HVS is sensitive to them; Orderly structure
Inspiration Theoretical Support: (Free-Energy Principle) According to the free-energy principle, orderly structures are easily to be predicted, and the HVS is sensitive to them; Disorderly structures possess abundant uncertainty, and the HVS is insensitive to them. Orderly structure Disorderly structures
Inspiration Theoretical Support: (Free-Energy Principle) According to the free-energy principle, orderly structures are easily to be predicted, and the HVS is sensitive to them; Disorderly structures possess abundant uncertainty, and the HVS is insensitive to them. Structural uncertainty determines the HVS sensitivity. Orderly structure Disorderly structures
Luminance Adaptation Luminance Adaptation JND Threshold JND Threshold Factors determine JND threshold : Contrast Masking Contrast Masking Structural Uncertainty Structural Uncertainty Inspiration
Luminance Adaptation Luminance Adaptation JND Threshold JND Threshold Factors determine JND threshold : Contrast Masking Contrast Masking Structural Uncertainty Structural Uncertainty How to calculate structural uncertainty? Inspiration
JND threshold estimation Structural Uncertainty: Represent the disorderly degree of the visual structure;
JND threshold estimation Structural Uncertainty: Represent the disorderly degree of the visual structure; Decompose disorderly content from orderly content; Disorderly content Acquisition Disorderly content Acquisition AR prediction Based decomposition AR prediction Based decomposition TheoryTechnology
JND threshold estimation Structural Uncertainty: Represent the disorderly degree of the visual structure; Decompose disorderly content from orderly content; Analyze the visual structure of the disorderly content; Disorderly content Acquisition Disorderly content Acquisition AR prediction Based decomposition AR prediction Based decomposition TheoryTechnology Visual Structure Characteristic Visual Structure Characteristic Local Binary Pattern (LBP) Local Binary Pattern (LBP)
JND threshold estimation Structural Uncertainty: Represent the disorderly degree of the visual structure; Decompose disorderly content from orderly content; Analyze the visual structure of the disorderly content; Calculate the structural uncertainty. Disorderly content Acquisition Disorderly content Acquisition AR prediction Based decomposition AR prediction Based decomposition TheoryTechnology Visual Structure Characteristic Visual Structure Characteristic Local Binary Pattern (LBP) Local Binary Pattern (LBP) Structural Uncertainty Structural Uncertainty Shannon Entropy Shannon Entropy
JND threshold estimation AR based disorderly content decomposition: AR based prediction:
JND threshold estimation AR based disorderly content decomposition: AR based prediction: The correlation between the central pixel and its i-th neighbor The correlation between the central pixel and its i-th neighbor
JND threshold estimation AR based disorderly content decomposition: AR based prediction: Disorderly content decomposition: The correlation between the central pixel and its i-th neighbor The correlation between the central pixel and its i-th neighbor
Input AR Pred. AR based disorderly content decomposition: JND threshold estimation
Input Orderly content Disorderly content AR Pred. AR based disorderly content decomposition: JND threshold estimation
Structure analysis on the disorderly content: Local binary pattern (LBP) is successful in structure description; The structure of a pixel is described as the correlations with its surround neighbors: By assigning the binomial factors, LBP form is acquired: JND threshold estimation
Structural uncertainty computation: JND threshold estimation Shannon Entropy is adopted; Structural uncertainty is computed as the entropy of LBP on the disorderly content; The LBP distribution on a local region centered at The LBP distribution on a local region centered at
JND threshold estimation Structural Uncertainty: Original Image Structural Uncertainty
Pattern masking: JND threshold estimation Contrast Masking Affection of Structural Uncertainty
Pattern masking: JND threshold estimation Contrast Masking Affection of Structural Uncertainty
JND threshold estimation Pattern masking : parameter setting for contrast masking function Linear Fitting The ground truth (the circle points) from subjective contrast masking testing
Pattern masking : parameter setting for structural uncertainty function The ground truth comes from subjective viewing testing JND threshold estimation
JND computation model JND threshold estimation Luminance Adaptation Luminance Adaptation JND Threshold JND Threshold Pattern Masking Pattern Masking Luminance Contrast Luminance Contrast Structural Uncertainty Structural Uncertainty
JND computation model JND threshold estimation Luminance Adaptation Luminance Adaptation JND Threshold JND Threshold Pattern Masking Pattern Masking Luminance Contrast Luminance Contrast Structural Uncertainty Structural Uncertainty
Experimental Results Pattern Masking VS. Contrast Masking Contrast MaskingPattern Masking Original
Experimental Results Pattern Masking VS. Contrast Masking Contrast MaskingPattern Masking Original
Experimental Results Pattern Masking VS. Contrast Masking Contrast MaskingPattern Masking Original
Comparison among JND models : Wu2013Proposed Yang2005 Liu2011 Original Experimental Results
Comparison among JND models : Wu2013Proposed Yang2005 Liu2011 Original Experimental Results
Comparing with the existing JND models (Positive values means the proposed JND outperforms the compared model) Experimental Results
Comparing with the existing JND models (Positive values means the proposed JND outperforms the compared model) Experimental Results Subjective quality comparison (38 subjects are invited)
Conclusion The shortcoming of the existing JND model is analyzed; We suggest structural uncertainty is another factor which determines JND; Structural uncertainty is defined and calculated; Novel JND model is proposed; Experimental results demonstrate the effectiveness;
Related Works J. Wu, W. Lin, G. Shi, etc., “Pattern Masking Estimation in Image with Structural Uncertainty,” IEEE TIP. J. Wu, G. Shi, W. Lin, etc., “Just Noticeable Difference Estimation For Images with Free-Energy Principle,” IEEE TMM. J. Wu, F. Qi, and G. Shi. “Self-Similarity Based Structural Regularity for Just Noticeable Difference Estimation,” JVCI. J. Wu, W. Lin, and G. Shi. “Visual Masking Estimation Based On Structural Uncertainty,” ISCAS2013.(Best student paper).