Vectorial vs Raster

Technical description

G1 is a deterministic, generator-agnostic screen for figures that were rasterised, recompressed, or assembled from sources of differing quality, rather than exported once from vector plotting software. A chart drawn as vector graphics and rendered at high resolution is uniformly sharp, carries no compression blocking, and has clean anti-aliased edges; a rasterised or multiply-compressed chart shows measurable, non-uniform degradation. The indicator converts the image to grayscale and sums three signals, resolution consistency, JPEG blocking, and edge quality, into a 0 to 5 score (capped). It requires the image to be at least 32 by 32 pixels.

How it works

The implementation is deterministic, runs at Layer 1, and uses OpenCV operators on the grayscale image.

Resolution (sharpness) consistency. The image is tiled into 16 by 16 blocks. For each block the variance of its Laplacian is computed as a local sharpness measure, and blocks whose Laplacian variance falls below 5.0 are discarded as background. Across the remaining active blocks, the coefficient of variation CV = standard deviation / mean of the sharpness measures is computed. A CV > 1.5 contributes 2.0 and a CV > 0.8 contributes 1.0, since a clean export is uniformly sharp while a spliced or mixed-resolution figure is not. Blocks more than two standard deviations from the mean are reported as anomalous regions, up to five per figure.

JPEG blocking artifacts. JPEG compresses in 8 by 8 blocks, so recompression leaves a periodic discontinuity at every eighth pixel boundary on both axes. The absolute gradient is taken along columns and along rows; for each axis the mean gradient at the 8-pixel boundary positions is divided by the mean gradient at the non-boundary positions to give a blocking ratio, and the two axes are averaged (an axis that is essentially flat is ignored). A combined ratio above 1.5 contributes min(2.0, (ratio − 1.0) × 1.0).

Edge anti-aliasing quality. Edges are found with a Canny detector (thresholds 50 and 150); figures with fewer than ten edge pixels are skipped. The Sobel gradient magnitude is sampled at the edge pixels, the 90th percentile of those magnitudes is taken as a high-gradient threshold, and the fraction of edge pixels whose gradient exceeds 70 percent of that threshold is the high-ratio. A high-ratio above 0.7, meaning most edges are abrupt rather than smoothly anti-aliased, contributes min(2.0, (high_ratio − 0.3) × 2.5), the signature of a rasterised rather than vector-rendered chart.

Aggregation. The three contributions are summed and reported as min(5.0, total). The metadata returns the resolution coefficient of variation, the number of blocks analysed, the JPEG-artifact and edge-quality scores, and the anomalous-block list.

Score thresholds

Score	Meaning
0 to 1	Uniform sharpness, no blocking, smoothly anti-aliased edges. Consistent with a single clean vector export.
2 to 3	One degradation signal present: uneven sharpness, visible JPEG blocking, or jagged edges.
4 to 5	Several signals together: non-uniform sharpness plus blocking plus rasterised edges. Consistent with a recompressed, rasterised, or multiply-sourced figure.

Why this matters

Manipulated and synthetic figures have become a measurable share of research-integrity cases: a 2025 review reported that roughly a third of flagged academic-misconduct cases now involve AI-altered images. The detection problem is hard in a specific way. The first benchmark for AI-generated scientific figures found that learned detectors transfer poorly across generators, overfit to the model they were trained on, and degrade under ordinary post-processing such as recompression and resizing. That is exactly the regime where a deterministic, generator-agnostic forensic signal earns its place: it does not try to recognise a particular generator's fingerprint but measures physical traces of how the image was produced and handled. Two such traces are robust. Sharpness that varies across a figure is the classic signature of splicing, since regions taken from different sources or rendered at different resolutions carry different local detail, and methods that expose splicing in scientific publications rely on exactly this kind of local inconsistency. JPEG blocking is the signature of recompression: the codec's 8 by 8 grid imprints a periodic discontinuity that a single clean export does not have. G1 reads both, and the anti-aliasing signal on top, to flag a figure whose pixels do not match the clean-export story its format implies.

Limitations

G1 measures degradation traces, so a high-quality figure rendered uniformly, whether genuinely vector-exported or generated cleanly, will not trigger the resolution or blocking signals; the indicator screens for inconsistency and recompression rather than for synthesis as such, which is the domain of the learned detectors it complements. The thresholds were calibrated against typical chart imagery and are directional rather than exact; line art and figures with legitimately sharp, deliberately hard edges can raise the edge-quality signal without manipulation. The check requires an image of at least 32 by 32 pixels and skips near-flat regions as background. Deeper compression and noise forensics, error-level analysis, frequency-domain fingerprints, noise-residual inconsistency, and double-JPEG (ghost) analysis, deliberately live in sibling image indicators, so G1 stays on the resolution, blocking, and edge axes to avoid duplicating them.

Theoretical background

G1 combines three established forensic ideas. The first is sharpness and noise inconsistency as a splicing cue: a composited image carries regions whose local detail statistics differ, and forensic methods, including recent work targeting scientific publications specifically, localise tampering by finding those inconsistent regions; G1 takes the coefficient of variation of per-block sharpness as a lightweight global version of that signal. The second is the JPEG block grid: because the codec quantises 8 by 8 blocks independently, recompression imprints a periodic discontinuity at the block boundaries, which compression-artifact detectors exploit to localise manipulation; G1 measures that periodicity on both axes. The third is the generalisation gap of learned synthetic-figure detectors, documented by the first scientific-figure benchmark, which motivates pairing them with deterministic, physically-grounded features that do not depend on having seen a generator before.

References

1. Hu Y, Zhao C, Mo C, Liu H, Li X. SciFigDetect: a benchmark for AI-generated scientific figure detection. arXiv preprint arXiv:2604.08211. 2026. https://arxiv.org/abs/2604.08211
2. Lin X, Tang W, Wang H, Liu Y, Ju Y, Wang S, Yu Z. Exposing image splicing traces in scientific publications via uncertainty-guided refinement. Patterns. 2024. DOI: 10.1016/j.patter.2024.101038
3. Kwon MJ, Nam SH, Yu IJ, Lee HK, Kim C. Learning JPEG compression artifacts for image manipulation detection and localization. International Journal of Computer Vision. 2022. https://arxiv.org/abs/2108.12947
4. From evidence to verdict: an agent-based forensic framework for AI-generated image detection. arXiv preprint arXiv:2511.00181. 2025. https://arxiv.org/abs/2511.00181