Wu et al (TIP)

Wu et al. (pattern complexity) · Pattern-aware JND with orientation diversity

Lineage: Pattern-aware Domain: Pixel Reference: Wu, J., Li, L., Dong, W., Shi, G., Lin, W., and Kuo, C.-C. J., "Enhanced just noticeable difference model for images with pattern complexity," IEEE Transactions on Image Processing, vol. 26, no. 6, pp. 2682–2693, June 2017.

This directory contains the OpenJND implementation of Wu et al.'s pattern-complexity JND model — a natural successor to texture-masking models for high-resolution natural imagery. The model captures the observation that two regions with identical contrast can mask very different amounts of distortion depending on whether their local patterns are regular (e.g. a brick wall) or irregular (e.g. crumpled fabric).

What the model does

The final JND map combines a luminance-adaptation branch with a visual-masking branch via the nonlinear additive masking model. The visual-masking branch is itself the dominance of two transducers — a classical luminance-contrast transducer and a new pattern-masking transducer driven by orientation diversity. Edge protection is applied to the pattern-masking branch only.

1. Luminance adaptation jnd_LA. A 5×5 weighted lowpass on the image gives the local background luminance bg; a dark-region adjuster lifts very-low-luminance pixels (min_lum = 32) to avoid runaway JND in shadows. The per-pixel threshold is the Chou-style piecewise formula

   bg_jnd(L) = { T₀ · (1 − √(L/127)) + 3      if L ≤ 127
               { γ · (L − 127) + 3              if L > 127

with T₀ = 17 and γ = 3/128, then scaled globally by α = 0.7:

   jnd_LA(x,y) = α · bg_jnd( adjusted_bg(x,y) )

2. Luminance contrast L_c. Computed as the local standard deviation over a 5×5 window (R = 2). It feeds two downstream transducers.

3. Luminance-contrast transducer jnd_LC. A classical Watson-style gain-control form:

   jnd_LC = (a1 · L_c^2.4) / (L_c² + a2²),    a1 = 1.84,  a2 = 26

4. Pattern complexity P_c. This is the model's distinctive contribution. For each pixel, an 8-neighbour ring (radius r = 1) plus the centre supplies 9 local gradient orientations. Orientations are computed from a 3×3 box-difference operator and quantised at a half-width of otr = 6° (giving 16 bins over [0°, 180°] plus one invalid bin for low-gradient pixels). P_c is the number of distinct orientation bins present in the ring (the L₀ norm of the orientation histogram). A 3×3 Gaussian (σ = 1) then smooths the complexity map.

5. Pattern-masking transducer. Combines local contrast with complexity:

   C_t    = (a3 · P_c^a4) / (P_c² + a5²),     a3 = 0.3,  a4 = 2.7,  a5 = 1
   jnd_PM = L_c · C_t

6. Edge protection. A Canny-based mask (adaptive threshold ≤ 0.8, disk-3 dilation, 5×5 Gaussian σ = 0.8) suppresses the pattern-masking budget along edges:

   jnd_PM_p = jnd_PM · edge_protect

7. Visual-masking dominance and NAMM combiner. The visual-masking branch is the stronger of the two transducers; the final JND combines it with luminance adaptation via Yang et al.'s NAMM:

   jnd_VM  = max( jnd_LC, jnd_PM_p )
   jnd_map = jnd_LA + jnd_VM − 0.3 · min( jnd_LA, jnd_VM )

The demo script also performs a JND-guided noise injection (adjuster = 0.7) to produce a contaminated image img_jnd for visual evaluation; this is a convenience for the demo, not part of the JND estimator proper.

Behaviour of the JND map

• Irregular-pattern regions receive a higher JND budget than regular-pattern regions of the same contrast, because P_c is higher there and the pattern transducer wins over the plain luminance-contrast transducer.
• Regular high-contrast edges are still protected — both because jnd_LC dominates there (the pattern-complexity branch is small near simple edges) and because edge_protect actively suppresses the pattern-masking term on detected edges.
• One of the best-performing pixel-domain bottom-up models on natural imagery, and in the OpenJND runtime analysis its Python port tends to be the fastest of the catalogue thanks to vectorised orientation statistics.

Directory layout

Wu et al (TIP)/
├── MATLAB/
│   ├── demo_pattern_complexity_based_JND_modeling.m   # entry-point demo
│   └── func_JND_modeling_pattern_complexity.m         # core algorithm
├── Python/
│   └── main.py
└── paper.pdf

Unified calling convention

INPUT  : grayscale image  (uint8 / float, H × W)
OUTPUT : 5-tuple (img_jnd, jnd_map, jnd_LA, jnd_VM, P_c)
         where jnd_map is the JND profile used by the OpenJND benchmarks
         and the other elements are intermediate diagnostics.

Minimal usage example

MATLAB

addpath('MATLAB');
img = imread('../test_data/lena.png');
if size(img, 3) == 3, img = rgb2gray(img); end
[img_jnd, jnd_map, jnd_LA, jnd_VM, P_c] = func_JND_modeling_pattern_complexity(img);
imshow(mat2gray(jnd_map));

Python

cd Python
pip install numpy scipy opencv-python matplotlib
python main.py

Programmatic call:

from main import func_JND_modeling_pattern_complexity
import cv2
img = cv2.imread('../test_data/lena.png', cv2.IMREAD_GRAYSCALE)
img_jnd, jnd_map, jnd_LA, jnd_VM, P_c = func_JND_modeling_pattern_complexity(img)

Default parameters

Parameter	Default	Meaning
T₀, γ (LA)	17, 3/128	Chou-style luminance adaptation curve
min_lum	32	Dark-region adjuster to keep jnd_LA finite in shadows
α (LA scale)	0.7	Global scale of jnd_LA
R (variance window)	2 → 5×5	Window for the local std-deviation that defines L_c
a1, a2 (jnd_LC)	1.84, 26	Constants of the luminance-contrast transducer
r (neighbour ring)	1 → 8 neighbours	Ring radius for orientation sampling
otr	6°	Half-width of an orientation bin (≈12° centre-to-centre, 16 bins + 1 invalid)
a3, a4, a5 (C_t)	0.3, 2.7, 1	Constants of the complexity transducer
Gaussian on P_c	3×3, σ = 1	Smoothing of the complexity map
Edge-protect Canny	adaptive, ≤ 0.8	Threshold from edge_h = 60 / max(edge_height), capped at 0.8
Edge-protect kernels	disk(3), 5×5 Gauss σ = 0.8	Dilation and smoothing of the edge mask
C (NAMM)	0.3	NAMM overlap-reduction factor
adjuster (noise demo)	0.7	Magnitude of the demo's JND-guided noise injection

All defaults reproduce the configuration used in the reference implementation.

Citation

@article{wu2017enhanced,
  title   = {Enhanced just noticeable difference model for images with pattern complexity},
  author  = {Wu, Jinjian and Li, Leida and Dong, Weisheng and Shi, Guangming and Lin, Weisi and Kuo, C.-C. Jay},
  journal = {IEEE Transactions on Image Processing},
  volume  = {26}, number = {6}, pages = {2682--2693}, year = {2017}
}