Semantically Adaptive JND Modeling with Object-Wise Feature Characterization, Context Inhibition and Cross-Object...

Area of Science:

• Computer Vision
• Human Visual Perception
• Signal Processing

Background:

• Existing Just Noticeable Difference (JND) models face performance bottlenecks due to reliance on low-level features.
• Current JND models inadequately capture the significant impact of high-level semantic information on perceptual attention and subjective video quality.

Purpose of the Study:

• To investigate the influence of heterogeneous semantic features on visual attention for optimizing JND models.
• To improve the efficiency and accuracy of JND models by integrating semantic understanding.

Main Methods:

• Analyzed visual attention responses to semantic features across object, context, and cross-object interactions.
• Quantified the coupling of visual features with Human Visual System (HVS) properties.
• Developed a semantic attention model incorporating object features, contextual complexity, and bias competition.
• Fused the semantic attention model with a spatial attention model using a weighting factor for a transform domain JND model.

Main Results:

• The proposed JND model demonstrates high consistency with HVS perception.
• Simulation results show the model is competitive with state-of-the-art JND models.
• The integration of semantic features significantly enhances JND prediction accuracy.

Conclusions:

• Semantic features are crucial for accurate JND modeling and video quality assessment.
• The proposed semantic attention-based JND model offers improved performance and efficiency.
• This work provides a pathway for developing more perceptually relevant video quality metrics.