photo_judgers/docs/research-brisque-niqe.md

# BRISQUE / NIQE — Classical No-Reference IQA

## Model Info

- **BRISQUE:** "No-reference Image Quality Assessment in the Spatial Domain" — Mittal et al., IEEE TIP 2012
- **NIQE:** "Making a 'Completely Blind' Image Quality Analyzer" — Mittal et al., IEEE SPL 2013
- **Python packages:**
  - `pip install brisque[opencv-python]` — Latest BRISQUE (supports opencv-contrib, opencv-headless)
  - `pip install pybrisque` — Alternative BRISQUE (older, from Bukalapak)
  - `pip install image-quality` — Contains `imquality.brisque` module
  - `pip install pyiqa` — IQA-PyTorch toolbox (includes BRISQUE + NIQE + 30+ other metrics)
  - `pip install nr-iqa` — Comprehensive library (BRISQUE + NIQE + PIQE)
- **License:** Academic (BRISQUE: BSD-like via LearnOpenCV; NIQE: academic use)
- **Dependencies:** OpenCV, NumPy, SciPy (BRISQUE); skvideo + NumPy (NIQE)

## Architecture

### BRISQUE (Blind/Referenceless Image Spatial Quality Evaluator)

- **Type:** Handcrafted features + Support Vector Regression (SVR)
- **Pipeline:**
  1. **MSCN (Mean Subtracted Contrast Normalized) coefficients** — decorrelates image from local brightness/contrast
  2. **GGD (Generalized Gaussian Distribution) fitting** to MSCN coefficients — extracts shape parameter alpha (tail heaviness)
  3. **Pairwise products** in 4 orientations (horizontal, vertical, 2 diagonals) — 16 features
  4. **AGGD (Asymmetric GGD) fitting** to pairwise products
  5. **Multi-scale analysis** at 2 scales (original + 0.5x downsampled) — 36-D feature vector total (18 per scale)
  6. **SVR prediction** maps 36-D features to quality score in [0, 100]
- **Training data:** TID2008 dataset (human-rated distorted images, MOS scores 0-100)
- **Score direction:** **Lower = better** (0 = perfect, 100 = worst quality)

### NIQE (Natural Image Quality Evaluator)

- **Type:** Handcrafted features + Multivariate Gaussian (MVG) model
- **Key difference from BRISQUE:** Completely opinion-unaware — does NOT need human-rated distorted images for training
- **Pipeline:**
  1. Extract 18-D NSS features from 96x96 patches (same as BRISQUE scale-1)
  2. Fit MVG model (mu_ref, Sigma_ref) to pristine images — the "naturalness" reference
  3. Fit MVG (mu_test, Sigma_test) to test image patches
  4. **Quality = Mahalanobis distance** between the two MVG models
- **Training data:** Only pristine (undistorted) images — no human ratings needed
- **Score direction:** **Lower = better** (0 = perfectly natural, higher = more distorted)

## Installation

```bash
# Option 1: Latest BRISQUE (recommended)
pip install brisque[opencv-python]

# Option 2: For servers/Docker (headless)
pip install brisque[opencv-python-headless]

# Option 3: Comprehensive NR-IQA library (BRISQUE + NIQE + PIQE)
pip install nr-iqa

# Option 4: IQA-PyTorch toolbox (BRISQUE + NIQE + 30+ other metrics)
pip install pyiqa
```

## Python Usage

```python
# --- BRISQUE via brisque package ---
from brisque import BRISQUE
import cv2

brisque = BRISQUE()
img = cv2.imread('photo.jpg')
score = brisque.score(img)  # Returns float in [0, 100]
print(f"BRISQUE: {score:.2f}")  # Lower = better

# --- BRISQUE via image-quality package ---
from imquality.brisque import BRISQUE
from PIL import Image

img = Image.open('photo.jpg')
score = BRISQUE().score(img)  # Returns float in [0, 100]

# --- NIQE via nr-iqa package ---
from nr_iqa import NIQE
import cv2

niqe = NIQE()
img = cv2.imread('photo.jpg', cv2.IMREAD_GRAYSCALE)
score = niqe.score(img)  # Returns float in [0, 100]
print(f"NIQE: {score:.4f}")  # Lower = better

# --- Both via IQA-PyTorch toolbox ---
import pyiqa
device = 'cuda' if torch.cuda.is_available() else 'cpu'

brisque = pyiqa.create_metric('brisque', device=device)
niqe = pyiqa.create_metric('niqe', device=device)

score_b = brisque('./photo.jpg')
score_n = niqe('./photo.jpg')

print(f"BRISQUE: {score_b:.2f}, NIQE: {score_n:.4f}")

# Batch processing
import glob
for path in glob.glob('./wedding-photos/*.jpg'):
    score = brisque(path)
```

## Score Interpretation for Wedding Photos

| BRISQUE Score | NIQE Score | Interpretation | Recommended Action |
|---------------|-----------|---------------|-------------------|
| 0-10 | 0-3 | Excellent technical quality | Auto-keep |
| 10-25 | 3-6 | Good technical quality | Keep |
| 25-40 | 6-10 | Acceptable, minor issues | Review |
| 40-60 | 10-15 | Poor technical quality | Likely reject |
| 60-100 | 15+ | Very poor (blurry, noisy, compressed) | Auto-reject |

**Recommended thresholds for wedding culling:**

- BRISQUE < 30 AND NIQE < 8 -> keep (good technical quality)
- BRISQUE > 60 OR NIQE > 15 -> reject (poor technical quality)
- In between -> manual review

## Performance Benchmarks

| Metric | SRCC vs Human | PLCC vs Human | Notes |
|--------|--------------|--------------|-------|
| BRISQUE (TID2008) | ~0.85 | ~0.87 | Trained on TID2008 |
| NIQE (KADID-10k) | ~0.83 | ~0.85 | Opinion-unaware, generalizes well |
| BRISQUE (live) | ~0.81 | ~0.83 | Live (no-reference) dataset |
| NIQE (live) | ~0.80 | ~0.82 | Live (no-reference) dataset |

## Hardware Requirements

- **CPU-only:** No GPU needed — purely classical (no neural networks)
- **Speed per image:** ~50-200ms per image on modern CPU (single-core)
- **Speed for 4,000 images:** ~3-15 minutes on CPU (single-core, depends on image size)
- **Multi-threading:** Can parallelize across CPU cores — ~1-3 minutes with 8+ cores
- **VRAM:** None required (BRISQUE/NIQE are pure NumPy/OpenCV)
- **Memory usage:** ~50-100 MB RAM for batch processing

## What It's Good At

- **No GPU needed:** Runs on any CPU, even low-end machines
- **Fast:** Good for quick technical quality screening — faster than any deep learning model
- **Detects blur and compression artifacts:** Excellent at detecting out-of-focus shots, motion blur, JPEG artifacts, noise
- **Simple:** Minimal dependencies (OpenCV + NumPy + SciPy)
- **Deterministic:** No randomness — same image always gets same score
- **Lightweight:** ~50 MB total install size, no model weights to download
- **Good complement to aesthetic models:** Catches technical rejects that aesthetic models miss
- **Works on any image type:** Not biased toward any particular visual style

## What It's Bad At

- **Purely technical:** Scores only technical quality (blur, noise, compression), not aesthetics at all
- **Poor on artistic/creative images:** Penalizes artistic choices — shallow depth of field, high contrast, film grain, vignette, intentional blur
- **Outdated methodology:** 2012-2013 techniques; inferior to modern deep learning approaches for complex distortions
- **Sensitive to intentional artistic effects:** May score a portrait with beautiful bokeh (shallow DOF) as "low quality"
- **Color-blind:** BRISQUE operates on luminance channel; NIQE on grayscale — ignores color information entirely
- **Not suitable as sole scorer:** Must be combined with an aesthetic model for wedding photo sorting
- **No multi-scale like MUSIQ:** Only processes at one scale (BRISQUE does 2-scale but simpler than MUSIQ's 3-scale)
- **Training data limitations:** BRISQUE trained on TID2008 (lab-distorted images); may not generalize well to real-world wedding photo distortions