Scratch-aware tagging for broadcast archives

Many archive collections contain images that look usable at a glance but carry evidence of age and handling. Fine vertical or diagonal scratches can cross faces, titles and scenery; emulsion damage, dirt and wear from repeated tape winding add other defects. Because a condition problem is usually absent from a catalogue entry, editors may rediscover it only after a clip has been selected, downloaded and prepared for transmission. Turning visible scratches into structured metadata changes that sequence of discovery.

The centre of ReCAP’s method for identifying and tagging archive material with visible scratches is the conversion of a subjective viewing complaint into time-specific, machine-readable evidence. A candidate scratch is detected within a frame or shot, linked to the source asset and assigned a confidence value that can be reviewed or combined with other analysis. The system is not presenting scratch detection as a restoration service. It provides a way to locate, classify and expose the condition of material before an editor commits to a particular excerpt.

That capability sits within ReCAP’s wider effort to support broadcast-quality content analysis. The EU-funded Real-time Content Analysis and Processing initiative addresses automatic metadata extraction, video-quality monitoring, recognition of faces and logos, and detection of duplicated content across production, live broadcasting and media-asset-management systems. Scratch tags can therefore travel through a wider set of editorial and technical services. They may inform a quality dashboard during a live feed, a preservation workflow during digitisation, or a search result months later when a producer needs an establishing shot from an older programme.

For Australian archives, the value is practical rather than academic. Public collections, commercial libraries and production teams often hold decades of broadcast and film material recorded in formats and technical environments that differ from today’s camera originals. A method that makes damage findable can reduce repeated manual inspections, reveal which copies of an asset are most damaged, and help decide whether restoration should happen before reuse. Automation supplies consistency and scale; trained archivists and editors still supply context, especially where a deliberate visual effect resembles a defect.

Why scratches need metadata

A scratch is an artefact of the recorded carrier or its imaging surface, not simply a thin bright shape. Film scratches may expose or distort emulsion, while tape or transfer problems can introduce vertical dropout-like lines. Dust, compression noise, subtitles and set geometry can look similar in one frame. ReCAP’s useful distinction is therefore between an isolated visual candidate and a defect supported by shape, persistence and context.

Metadata maps the condition to an exact place in the asset. Rather than marking an entire reel as damaged, the record can point to a start and end timecode, indicate the frame region, name the suspected artefact and retain a confidence score. Search, filters and quality reports can then use those fields without changing the moving image itself. This preserves the source while making risk visible to systems and people who never open the file.

The consortium’s wider technical objectives are set out in the ReCAP project overview, including its work on real-time analysis, automatic metadata and quality monitoring. For archive teams, the important point is interoperability: scratch information should sit beside source identifiers, ingest history and technical metadata, not in a separate report that an editor may never encounter.

Inside the detection process

At ingest, decoded frames provide the evidence surface for analysis. The process looks for abrupt pixel transitions and narrow linear features that may indicate damage, then examines neighbouring frames to see whether the feature behaves like a physical artefact. A line that persists in a consistent image position can behave differently from scenery moving through the shot or a momentary encoder error. Temporal context reduces the chance that normal picture content is promoted to a damage label.

Classification then combines the spatial evidence with information from the sequence. Length, orientation, contrast, recurrence and relation to other frames can all contribute to a score. Thresholds matter: a low-confidence candidate can be retained as a review item, while a high-confidence event can trigger a stronger warning. This is more useful than a binary “damaged” or “undamaged” decision, particularly for mixed material containing both clean inserts and worn transfers.

Outputs must also remain robust across versions. The same clip may exist as a camera master, a broadcast recording and several transcodes. Compression can soften a scratch, sharpening can make noise prominent, and frame-rate conversion can duplicate or remove evidence. A traceable link between detection, model version and the exact file analysed prevents apparently conflicting results from being treated as new findings.

Tagging that survives cataloguing

Tagging begins with a compact, controlled vocabulary: scratch or surface artefact, severity, location, time span, confidence and evidence reference. A catalogue record can hold several events for one asset while keeping the human-readable summary separate from the machine-readable measurements. Shot-level tagging is often practical for editors, whereas frame-level anchors support close inspection and selective restoration.

Provenance matters as much as the label. The record should identify the source identifier, the analysed derivative, the timecode convention and whether the finding came from automatic detection, manual review or both. Corrections should be retained as new observations rather than silently overwriting the first result. This creates an audit trail when a broadcaster assesses suitability, an owner checks rights, or a preservation team compares multiple copies.

Tags should describe condition without pretending that technical damage determines editorial value. A scratched interview may still be the only record of a significant event, while a pristine clip may contain rights restrictions or inaccurate metadata. Scratch detection complements restoration planning and editorial judgement; it does not replace either. Its strongest role is to put a qualified alert in front of the right operator before reuse.

From live monitoring to archive ingest

ReCAP’s real-time framework supports two distinct operating patterns. In live production, a quality event can be surfaced while a feed is on air, giving a vision mixer or operator a chance to compare sources or switch away from an impaired feed. The alert is diagnostic: it identifies a probable artefact and its timing rather than explaining the cause or correcting the signal.

Archive ingest usually works differently. Analysis runs on preserved files, and results become part of a retrospective quality profile. Batch throughput matters because institutions can hold thousands of hours, but so does accuracy around sparse events in long programmes. Duplicate detection adds another layer: if the same scene appears in a clean master and a worn dub, a system can connect the copies and direct reviewers to the better source.

Different media introduce different baselines. A scan of scratched film, an analogue tape recording and a modern digital master should not be judged by identical expectations. Configurable thresholds and clear provenance make those differences visible. Integrated quality monitoring can then report both technical faults and content metadata through one workflow, reducing the gap between a preservation technician’s diagnosis and an editor’s search.

Keeping people in the decision loop

Automation is strongest when its boundaries are explicit. Reviewers need to see the affected frame, the candidate region, the confidence value and nearby events. A thumbnail alone may hide a hairline scratch, while a short sequence makes it easier to distinguish a defect from movement. Confirmed examples can improve thresholding and, where permitted, support future training without changing the historical status of the original finding.

False positives still have a cost. Repeated warnings for subtitles, stage lighting or intentional film effects can train staff to ignore the tool. False negatives are costly in a different way: clean-looking output may carry unnoticed damage. Sampling, priority rules and role-based review help balance those risks, especially when large holdings are processed faster than an archive can inspect them frame by frame.

For teams that already use Telegram for production coordination, Telegram Plus can organise clip-review threads around the affected timecode and a reviewer’s decision. The chat is not the analytical system of record; it is the route by which a concise alert reaches a producer, archivist or quality-control officer. Keeping the authoritative finding in the asset-management platform preserves governance while making collaboration faster.

Australian workflows in focus

Australian use cases span a distinctive local market. Sydney and Melbourne are major centres for television production, post-production and commercial media-asset management, while Canberra is home to the National Film and Sound Archive of Australia. The ABC and SBS also maintain substantial content libraries. These institutions deal with different access models, but all need reliable descriptions of what survives in their holdings.

The scale of Australian production makes condition data useful across several stages. Screen Australia-supported productions may have been filmed in coastal locations, the outback or dense urban environments, creating difficult visual backgrounds for an automated detector. Bush footage can contain fine branches, bright horizons and dust; studio material can contain lighting rigs, graphics and deliberate scratches. Context-aware analysis helps separate those subjects from damage.

Much legacy material was produced or captured in the PAL Standard Definition era and then transferred through analogue tape, broadcast recording and film scanning. The resulting archive can contain generations of noise and wear. Tagging a scratch at the point where it is actually visible keeps a later transfer from inheriting an unsupported assumption: the condition of one copy does not automatically describe every derivative.

Climate adds another local layer. Heat and dust can accelerate deterioration of physical media in less carefully controlled facilities, while humidity affects storage conditions near the coast. These environmental risks are not visible scratches, and should not be conflated with them, yet they make early condition monitoring valuable. Searchable scratch metadata can identify the footage most in need of handling review, preservation action or a better surviving copy.

What operators should remember

The method works best when detection, metadata and review are treated as one chain. A candidate must be tied to a particular frame and file; the tag must retain its confidence and provenance; and the resulting warning must reach someone able to judge editorial and preservation consequences. Architecture matters, but clear operating rules determine whether the information becomes useful.

The point to remember is simple: once visible scratches become timed, confidence-scored metadata, an archive can find, flag and responsibly reuse damaged material without mistaking every bright line for a defect.