Abstract
The just noticeable difference (JND) models in pixel domain are generally composed of luminance adaptation (LA) and contrast masking (CM), which takes edge masking (EM) and texture masking (TM) into consideration. However, in existing pixel-wise JND models, CM is not evaluated appropriately since they overestimate the masking effect of regular oriented texture regions and neglect the visual attention characteristic of human eyes for the real image. In this work, a novel JND model in pixel domain is proposed, where orderly texture masking (OTM) for regular texture areas and disorderly texture masking (DTM) for complex texture areas are presented based on the orientation complexity. Meanwhile, the visual saliency is set as the weighting factor and is incorporated into CM evaluation to enhance JND thresholds. Experimental results indicate that compared with existing relevant JND profiles, the proposed JND model tolerates more distortion in the same perceptual quality, and brings better visual perception in the same level of the injected JND-noise energy.
Introduction
The Human Visual System (HVS) possesses limited sensitivity, meaning not every change in an image is perceivable. In image processing, the Just Noticeable Difference (JND) threshold defines the maximum distortion a pixel can sustain before a human observer notices a change. Accurately modeling JND is vital for applications like image compression, watermarking, and quality assessment.
Traditional pixel-wise JND models primarily consider Luminance Adaptation (LA)—the HVS's varying sensitivity to brightness—and Contrast Masking (CM)—the phenomenon where textures and edges "hide" noise. However, these models often overestimate the masking effect in regular, repeating textures and fail to account for visual saliency, the cognitive process where the eye prioritizes specific "interesting" regions.
Proposed Methodology
The proposed model refines the calculation of Contrast Masking by decomposing the image and integrating visual attention.
Image Decomposition and Texture Analysis
Using a Relative Total Variation (RTV) model, the input image is split into a structural image (containing edges and smooth areas) and a textural image (containing fine details).
The textural image is further analyzed for "orientation complexity" to distinguish between:
- Orderly Texture Masking (OTM): Applied to regions with regular, predictable patterns where the HVS is more sensitive to irregularities.
- Disorderly Texture Masking (DTM): Applied to complex, chaotic areas (e.g., grass, fur) that provide stronger masking effects.
Visual Saliency and Final Integration
A bottom-up saliency map is generated to identify focal points. If a region is highly salient, the model applies a weighting factor to reduce the JND threshold, acknowledging that viewers are more likely to spot artifacts in these areas.
Finally, the LA and the refined CM (combining EM, OTM, DTM, and saliency) are merged using a Nonlinear Additivity Model for Masking (NAMM) to produce the final JND map.
Experimental Results
The model was evaluated using 12 standard test images ($512 \times 512$ resolution) and compared against state-of-the-art models by Liu et al. and Wu et al.
Objective Performance
The proposed model demonstrated superior noise redundancy. It achieved a lower Peak Signal-to-Noise Ratio (PSNR) while maintaining equivalent perceptual quality (SSIM). Specifically, it successfully hid 0.86 dB more noise than Liu’s model and 1.19 dB more than Wu’s model.
Subjective Evaluation
In subjective tests involving 20 human assessors, the proposed model consistently provided better visual quality. By intelligently placing noise in "disorderly" and non-salient regions, it avoided the visible artifacts often seen in regular textures or focal points when using older models.
Conclusion
By distinguishing between orderly and disorderly textures and incorporating visual saliency, this novel JND model more accurately reflects human perception. It allows for higher levels of data hiding and compression without compromising visual quality, paving the way for more efficient perceptual-oriented image and video processing.
Summary updated