Colour-aware analysis for non-standard video

ReCAP handles video with non-standard color primaries and gamut by treating colour as a measured property rather than a cosmetic filter. Cameras and post-production tools often record wide-gamut or logarithmic images, while files may carry tags that do not match the pixels. For Australian broadcast and streaming teams, the consequence extends beyond appearance: confidence scores, logo matches, duplicate-content checks and quality reports can become unreliable if analysis and delivery interpret the same image differently.

The European Union-funded Real-time Content Analysis and Processing initiative approaches this through metadata extraction, pixel-level inspection and traceable transformations. Across the consortium, its technical goals and demonstrators cover video-quality monitoring, face and logo recognition, and duplicate-content detection for production, live broadcasting and media asset management. The key design choice is to preserve source characteristics while creating comparable working representations for automated analysis.

Why non-standard colour reaches the workflow

Australian free-to-air HD has a familiar BT.709 reference, but production does not live there. Camera manufacturers offer wide-gamut modes, grading systems may use DCI-P3, and HDR deliverables can use BT.2020 with a perceptual quantiser or hybrid log-gamma transfer function. A team in Melbourne or Sydney may therefore receive a camera original, a DCI-P3 master, a BT.709 broadcast rendition and a BT.2020 streaming file for the same programme.

Field production adds variation. A Perth or Brisbane crew may record under strong daylight using a camera configured for a manufacturer-specific gamut or log curve. Editors then transform that image, but older files can arrive with missing side data, flattened graphics, or metadata copied from a template rather than derived from the actual render. In live workflows, a temporary feed can reach the gallery before the intended grade exists.

Legacy Australian material introduces another case. Four-by-three broadcasts and digital transfers often use BT.601-family conventions, while archive restoration can introduce grading changes. These files are not automatically defective. They need an explicit interpretation, especially when an automated service compares them with modern BT.709 or BT.2020 material.

Primary coordinates, gamut and transfer curves

Primaries specify where red, green and blue sit in colorimetry; together with a white point, they define the gamut triangle. A wider triangle can represent more saturated colours, but it does not automatically guarantee smoother tones or more recorded detail. An RGB working image may contain values outside the display gamut, and the display transform must map those values without unintended clipping.

The transfer function is a separate variable. Gamma, logarithmic encoding, PQ and HLG describe how scene or display values are represented numerically. Matrix coefficients determine how luminance and chrominance are combined. A file can therefore contain wide-gamut primaries with a conventional transfer curve, or BT.2020 metadata around content that was actually finished for BT.709. Treating the labels as one undifferentiated colour space removes information the analyser needs.

Full-range and limited-range signalling matter too. A 0–255 range may be compressed into narrower video levels, or a file may retain full-range flags after transcoding. These conditions can change numerical measurements without changing the intended colours. ReCAP’s metadata-first approach is useful here: it keeps primaries, transfer characteristics, matrix coefficients and range interpretation as separate facts rather than collapsing them into a single preset.

ReCAP’s colour-aware processing path

At ingest, ReCAP can extract the colour tags carried in the container and codec, then compare them with the file’s actual content. That comparison catches obvious mismatches, such as BT.709 flags on an image whose saturated colours clearly sit beyond the BT.709 gamut. It also creates an audit trail showing whether a later analysis used the declared colour interpretation or an overridden one.

Pixel-level checks add evidence to the tags. The analyser can calculate chromaticity under candidate primaries, look for extreme red, green or blue clipping, and assess whether neutral objects remain neutral after conversion. These are indicators, not proof of an error: a saturated Australian Football League jumper, a bright cricket boundary, a night-time city sign or a company logo can legitimately exceed a display gamut. Content context and uncertainty therefore belong in the report.

For comparable processing, the next step is a colour-managed transformation into a declared working representation. The source stays intact; a normalised copy can be linearised, adapted to a common white point and encoded for analysis. This avoids treating display-ready values as scene values. It also gives later modules—quality metrics, perceptual hashing and recognition—a stable basis on which to compare footage recorded with different cameras or grades.

The results should remain traceable even when the original gamut cannot be restored exactly. The report can record the source primaries, the target representation, the transform and any clipping. This matters for ReCAP’s detection functions: a face or logo found on a BT.709 proxy must still map back to the corresponding region in the original, while a perceptual duplicate match should tolerate colour changes without confusing them with the same content. Transformed evidence and original evidence should never be silently exchanged.

Preserving quality during analysis and compression

Conversion itself can degrade a signal. Out-of-gamut values must be mapped into a smaller triangle, while scene values need a suitable tone curve for the intended display. A colour-managed transform can make those choices explicit, but it cannot recreate detail that the camera or an earlier codec has already destroyed. ReCAP can expose clipped pixels and spatial colour excursions as quality indicators, alongside luma, temporal variation, block artefacts and Delta E comparisons in an appropriate reference space.

Compression is another variable. A separate review of image-quality compression tuning can help teams test bitrate and image fidelity without confusing codec degradation with a colour-management error. ReCAP should attribute artefacts to their source where possible: ringing around a logo is not the same defect as a cyan shift across a shot, and a low-bitrate proxy should not become the colour reference for the mezzanine master.

A sensible pipeline retains the acquisition or mezzanine file as the evidentiary source, generates a proxy for routine processing, and applies display-referred metrics to material intended for broadcast delivery. Scene-referred analysis can run separately for production decisions. Comparing the same frames before and after conversion also reveals whether a transform has introduced banding, shifted skin tones or crushed highlights. That separation is particularly useful for automated checks, where a uniform but unwanted cast can be as harmful as visible clipping.

Licensed content, recognition and Australian safeguards

Colour handling and rights handling meet in licensed or sponsored footage. The project’s account of automated watermark detection shows how an overlay can be located in video and checked over time. A coloured strip, translucent bug or station logo may occupy bright or saturated pixels, so its presence must not be mistaken for an out-of-gamut error. Detection is stronger when it accounts for the source’s declared colour interpretation.

A watermark or recognisable company logo is evidence to review, not proof that a broadcaster owns a clip. Rights teams still need the relevant licence, contract or chain-of-title record. In Australia, that record sits alongside copyright and licensing obligations under the Copyright Act 1968. A ReCAP-based workflow can connect an automated finding to source timecode and a confidence score; it should not make a final legal determination about ownership, consent or permitted reuse.

Face recognition brings a different boundary. Facial images may be personal information, and biometric information used for automated identification is subject to additional considerations under the Privacy Act 1988 and the Australian Privacy Principles. Matching a face or logo should not be treated as evidence of consent. If an Australian online service also classifies material, the Online Safety Act 2021 may add duties; colour analysis itself is not a substitute for those compliance processes.

Australian testing conditions and service realities

Australia’s geography makes distributed review routine. Production teams in Brisbane and Melbourne may work with footage recorded in Perth, while Canberra-based editors coordinate with colleagues in other states and time zones. Sydney and Melbourne share the same eastern time, whereas Brisbane is ahead of them and Perth is ahead again. A remote approval late in the Australian evening can coincide with a live operation elsewhere, leaving little room for an unnoticed colour mismatch.

The local viewing market also shapes the test set. Many households watch sport and news on television, often on a calibrated broadcast monitor in production but on a mixture of older and newer consumer displays at home. Mobile viewing adds smaller screens, different browser colour management and varying brightness. A vivid daytime football fixture in Adelaide needs different expectations from a dimly lit drama scene, even when both use the same mastering specification.

Delivery infrastructure adds another constraint. National Broadband Network plans, fixed-wireless services and 5G connections have different performance characteristics, and viewers on regional or mobile links may receive an aggressively compressed rendition. ReCAP should therefore test both the visual signal and the transport path. Distinguishing a colour-management fault from a transcoding artefact helps teams decide whether to revisit the grade, the encode settings or the delivery configuration instead of repeatedly altering pixels that were already correct.

A practical quality-control protocol

A repeatable workflow should preserve evidence, describe transformations and separate factual signals from inference. For teams evaluating ReCAP, the following controls provide a practical baseline:

The next concrete step is to assemble one ten-second representative clip from an Australian production chain, preserve its original colour metadata, and run it through the ReCAP demonstrator against a BT.709 reference before approving any automatic conversion or alert.