BRISQUE Extractor¶

BRISQUE (Blind/Referenceless Image Spatial Quality Evaluator) is a no-reference, opinion-aware blind image-quality metric (Mittal, Moorthy & Bovik, IEEE Transactions on Image Processing 21(12):4695-4708, 2012). It scores a single picture from its spatial natural-scene statistics (NSS) and a trained support-vector regressor (SVR), without a reference frame. The fork ships it as a libvmaf feature extractor named brisque.

Because it is a no-reference metric, BRISQUE scores only the distorted picture; the reference frame and the 90°-rotated inputs are ignored, mirroring the CAMBI / NIQE no-reference posture. Lower score means better perceptual quality.

Output¶

Usage¶

vmaf \
    --reference dist.yuv \
    --distorted dist.yuv \
    --width 1920 --height 1080 --pixel_format 420 --bitdepth 8 \
    --feature brisque \
    --output score.json

BRISQUE ignores the reference, so you may pass the same file for --reference and --distorted (or any reference of matching dimensions). The per-frame JSON metric key is brisque; pooled values appear under the same key in pooled_metrics.

Options¶

Pass it like any feature option (= separates the feature name from its options, : separates options):

vmaf … --feature brisque=model_path=/path/to/brisque_live.model …

Model loading¶

The bundled LIVE-lab allmodel is embedded into the libvmaf binary at build time. core/src/meson.build runs an xxd -i custom_target over model/other_models/brisque_live.model (the ~343 KB native libsvm text model) to generate brisque_live.model.c in the build directory; that translation unit defines the src_brisque_live_model[] / src_brisque_live_model_len symbols declared in core/src/feature/brisque_model.h. This is the identical pattern libvmaf uses to bake in its own SVR JSON models, so BRISQUE carries no runtime model-file dependency and nothing needs to be installed alongside the binary.

At init() the extractor loads the model in this order:

1. If the model_path option is set, svm_load_model(model_path) reads that on-disk file (used to override the bundled model, or as the only loader when built_in_models is disabled).
2. Otherwise the embedded buffer is parsed with svm_parse_model_from_buffer(src_brisque_live_model, src_brisque_live_model_len).

A missing/invalid path, or a build with neither the embed nor a model_path, fails init() with -EINVAL and a descriptive log line.

The model bytes are not committed as a C source file: the byte array expands to ~2.1 MB, which exceeds the repository's 1 MB large-file gate. Only the ~343 KB binary model and the tiny declaration header live in the tree; the big array exists solely in the build directory.

Algorithm¶

Per frame, on the distorted luma plane only (all double precision), replicating the gregfreeman MATLAB pipeline that trained the model (brisque_feature.m):

1. Read the luma into a [0, 255] double working buffer (8-bit copied directly; higher bit depths scaled value * 255 / (2^bpc - 1)).
2. For each of two scales (full resolution, then ½-resolution):
3. MSCN — mu = filter2(gauss7x7, img), sigma = sqrt(|filter2(gauss7x7, img²) − mu²|), mscn = (img − mu) / (sigma + 1). The 7×7 Gaussian uses sigma = 7/6 (separable, unit-volume, zero boundary — MATLAB filter2('same')).
4. GGD fit of the MSCN field → push {alpha, sigma²} (2 features).
5. AGGD fit of each of the 4 paired products (shifts (0,1),(1,0),(1,1),(-1,1), wraparound roll) → push {alpha, eta, leftstd², rightstd²} (16 features).
6. Downsample img by ½ (MATLAB antialiased bicubic, Catmull-Rom a=-0.5).
7. The 36-D vector is [scale1 f1..f18, scale2 f1..f18].
8. Range-scale each feature to [-1, 1] via xs = -1 + 2·(f-min)/(max-min) (no clamp).
9. Predict: score = svm_predict(allmodel, x) over the embedded EPSILON_SVR RBF model (gamma 0.05, total_sv 770, rho -155.845).

The gamma-ratio tables for the GGD/AGGD shape fits are precomputed once at init() over the grid gam = 0.2 : 0.001 : 10 (9801 entries): r_ggd(g) = Γ(1/g)Γ(3/g)/Γ(2/g)², r_aggd(g) = Γ(2/g)²/(Γ(1/g)Γ(3/g)).

Load-bearing fidelity choices¶

The fork follows the MATLAB pipeline that trained the model, NOT the widely-copied krshrimali C++ port, which diverges on three points the model was never trained with:

• GGD (not AGGD) for the MSCN field (features f1, f2). krshrimali fits the MSCN field with AGGD — a bug versus both the paper (Table I) and the trained model. The model expects GGD.
• Gaussian sigma = 7/6 (not the truncated 1.166 in the C++ port).
• MATLAB antialiased bicubic ½-downscale (not OpenCV INTER_CUBIC).

Range arrays¶

The min_[36] / max_[36] range-normalization arrays are baked into core/src/feature/brisque.c and come from the inline arrays in the reference computescore.cpp (the array the trained model expects, and the only one the reference prediction path reads). They are not the separate allrange file in the same upstream repo, which is a different, prediction-inconsistent set the code never reads — substituting it would corrupt every score (an RBF gamma=0.05 perturbation over a 0.71 scaled-space shift is score-destroying, not a 1e-4 nudge). See ADR-1115.

Inputs and backends¶

• Bit depth: 8 / 10 / 12 / 16, scaled to the [0, 255] SDR working range.
• Pixel formats: all (luma-only; chroma is ignored, so 400P is fine).
• Backends: CPU scalar only. GPU twins (CUDA/SYCL/HIP) are out of scope; a future twin must keep the gamma argmin and RBF kernel sum in fp64 and is not expected to be bit-exact (per the fork's GPU-parity posture).

Limitations¶

• SDR luma only. BRISQUE is trained on 8-bit SDR LIVE images. PQ / HLG HDR is out of scope — no HDR-trained BRISQUE model exists. On >8-bpc input the extractor emits a one-time warning and scores the content as SDR; the score is not meaningful for HDR transfer characteristics.
• Near-zero AGGD sign sensitivity. The AGGD paired-product fit buckets samples by a strict sign cutoff (x<0 / x>0, zeros excluded). On heavily-compressed near-flat content a large fraction of paired products sit within ~1e-11 of zero, so the bucket assignment — and hence eta and the score — is sensitive to floating-point summation order (~0.1 score units). This is inherent to BRISQUE. Correctness is validated on stable natural content (cameraman: C extractor -13.70844 vs an independent MATLAB-faithful oracle -13.70840, ~5e-5); the in-tree fixture snapshot (testdata/scores_cpu_brisque.json) is the extractor's own deterministic output rather than a places=4 cross-assert.
• Minimum frame size: width and height each ≥ 7 (the 7×7 window must fit).

Updating the model and regenerating the score snapshot¶

There is no header to regenerate: the model is embedded straight from the vendored binary at build time. To swap in a different/retrained model, replace model/other_models/brisque_live.model (keep the libsvm text format) and rebuild — Meson re-runs xxd automatically. Update the model card (model/brisque_live_card.md), the NOTICE (model/other_models/NOTICE-brisque), and the score snapshot below to match.

To refresh the end-to-end snapshot after an intentional kernel change, run the extractor on frame 12 of the fixture and record the score:

core/build-cpu/tools/vmaf \
    --reference testdata/dis_576x324_48f.yuv \
    --distorted testdata/dis_576x324_48f.yuv \
    --width 576 --height 324 --pixel_format 420 --bitdepth 8 \
    --feature brisque --json --output /tmp/brisque.json
# update testdata/scores_cpu_brisque.json "score" and core/test/test_brisque.c

Correctness test¶

core/test/test_brisque.c (meson test_brisque, suite fast) asserts:

• gamma-table anchors GGD(2)=1.5707963, AGGD(1)=0.5, AGGD(2)=2/π;
• the 7×7 Gaussian window (unit sum, symmetric);
• GGD / AGGD fits of [-2,-1,0,1,2,3] (with zeros excluded from the AGGD buckets — right² = 14/3, not the NIQE zero-bucketing);
• a symmetric AGGD giving eta == 0 exactly and an all-zero flat-patch NaN guard;
• the MATLAB-imresize bicubic coefficients (normalized to 1, odd-dimension safe);
• the range-scale anchors (min→-1, max→+1, mid→0);
• an end-to-end score snapshot through the public extractor-context API;
• an odd-dimension regression (577×325) asserting a finite score and no bicubic-table overflow.

See also¶

• ADR-1115 — design + the model redistribution exception.
• docs/research/1101-brisque-nr-metric.md — research digest (constants, oracle, instability analysis).
• model/brisque_live_card.md — model card + required citation.
• NIQE — the sibling no-reference, opinion-unaware metric.