How ReCAP identifies and flags repeated ads across separate feeds

Commercials rarely stay confined to the channel that first carried them. A campaign can appear in a broadcaster’s linear feed, a catch-up replay, a regional variation and a separate recording, sometimes within minutes of one another. For Australian media teams, this creates a recurring operational question: has the same commercial simply been reused, or does a new placement contain something genuinely different?

ReCAP is the EU-funded Real-time Content Analysis and Processing initiative, a consortium project developing automated tools for broadcast-quality video analysis. Its work spans metadata extraction, quality monitoring, face and logo recognition, and the identification of duplicated material. That combination matters because a repeated ad is more than a title match: it may carry a changed offer, a different legal version or a new placement around it. The project also supports production and asset-management workflows, where a detected copy can be connected to catalogue metadata rather than found only through an editor’s memory.

The technical task has two layers. An exact duplicate can be matched directly, while a near duplicate needs tolerances for transcoding, cropping, compression, overlays and editorial edits. Binary file comparison fails on the first transcode. Perceptual video fingerprints, audio fingerprints and contextual cues can remain stable even when the encoded file changes.

Understanding how ReCAP identifies and flags repeated commercial content across separate feeds is therefore more useful than a simple binary comparison. The useful outcome is an explainable alert rather than a bare yes-or-no result. A strong flag connects the two feed locations, gives a confidence level and shows which visual, audio or textual evidence supported the match. For teams working with live advertising, that analysis may need to happen quickly enough to influence a broadcast decision, while archive searches can operate at a slower pace.

Why repeated commercials create operational problems

A “feed” may be a live television stream, an archived programme, a broadcaster app rendition or a file delivered to a playout system. The same commercial can legitimately appear in several of them because networks syndicate campaigns, sell national and regional inventory, and republish programmes. A detector must therefore recognise reuse without treating every second appearance as an error. The operational risk runs in both directions: missed copies waste rights and sales effort, while false positives interrupt legitimate scheduling.

A reliable decision draws on complementary signals:

Signal What it compares Handles feed variation Main contribution
Audio fingerprint Rhythm, dialogue, music and acoustic sequence Compression, volume changes and small picture edits Finds a familiar sound bed or full audio track
Visual fingerprint Key frames, composition, motion and shot order Re-encoding, resizing and modest cropping Links the same imagery across different renditions
Text and logo cues On-screen copy, brand marks and product names Changes to unrelated background imagery Connects localised or edited versions
Sequence alignment The order and duration of matching shots Different breaks and surrounding programmes Separates a real match from a generic resemblance

No one signal is dependable alone. Two ads may share a network ident or generic jingle without being the same campaign, while a new end card can hide an otherwise identical body. ReCAP’s value comes from combining evidence and preserving the reasons behind each decision. A confidence band is also more informative than a universal threshold because feeds can use different codecs, resolutions and capture paths.

The signals behind duplicate detection

The process begins when a feed is ingested. Frames and audio are decoded, time codes are normalised and the stream is segmented around shot changes, scene transitions, black frames and obvious breaks. This prevents the system from comparing entire multi-hour recordings and lets it search at the level of a commercial, bumper or other short item. For live feeds, the same indexes can be updated as segments arrive, allowing a spot seen on one channel to become a reference for the next.

From each candidate segment, ReCAP can derive a visual fingerprint and an audio fingerprint. The visual representation describes salient frames and their sequence; the audio representation follows the timing and spectral shape of speech, music and effects. Both can tolerate the loss of information introduced by a new delivery chain. Candidate fingerprints are then matched against an indexed reference set, including material previously seen in another feed.

Metadata and recognised elements strengthen the match. Extracted text may reveal the advertiser, product or campaign wording, while logo recognition identifies recurring brand marks. Face recognition can be useful in specialised catalogue or presenter workflows, but it is not the main basis for detecting a repeated commercial. Broadcast-quality checks also matter: dropped frames, severe compression or broken audio can reduce confidence and should be reported rather than silently treated as a content difference.

Matching and flagging across separate feeds

During ingestion, useful references are stored in a searchable fingerprint index. When a new segment arrives, ReCAP retrieves close visual and audio candidates, then tests them against the original occurrence rather than relying on filename or metadata similarity. Within the consortium, Nablet media analytics contributes expertise in analysing media streams and turning detected patterns into operational information.

The comparison follows the structure of the advertisement. A whole spot may align almost frame for frame, while a cutdown might contain only the middle shots, or a version may add an Australian offer at the end. Temporal alignment lets the system ask whether the shared material occurs in the same order and proportion. This step prevents a short shared legal line from outweighing the rest of the creative. Combined scoring can then classify the relationship as exact, edited, localised or uncertain, even when the feeds arrived through unrelated systems.

A generated flag should carry enough context for action. It can include the source feed, channel, start and end time, matched reference, confidence score, evidence types, detected creative version and current review state. Low-confidence matches remain visible for inspection rather than being promoted automatically. A review queue can group occurrences into one campaign cluster while preserving each individual placement, allowing a producer to confirm a scheduled campaign while an archive or rights team investigates an unexpected copy.

Australian broadcast realities and governance

Australian viewing habits make cross-feed reuse especially visible. Audiences in Sydney, Melbourne and Brisbane often move between live linear channels, free-to-air catch-up services, broadcaster apps, subscription streaming and recordings. The same commercial can therefore be encountered on television, watched later through a catch-up service and stored again in a personal library, even when it was broadcast once.

The local market adds variation. Australia combines commercial free-to-air networks, public broadcasters, regional affiliates and streaming services. An advertiser may buy national inventory while adapting a commercial for a state or regional audience, and schedules differ between eastern and western time zones. Local offers and Australian disclaimers can also change the duration of a break and the order in which creatives appear. These differences produce legitimate alternate versions and varied placement times, which makes commercial fingerprints more reliable than simplistic episode or schedule matching.

Legal context also needs care. Copyright and communication of protected material may engage the Copyright Act 1968, but a duplicate-content alert is evidence for review, not a finding of infringement. If facial analysis is combined with identifiable records, the Privacy Act 1988 and Australian Privacy Principles may also apply. Operational teams should minimise unnecessary personal data and keep content-matching decisions separate from any legal determination. Access controls, retention limits and documented review procedures can help organisations meet those obligations when video-derived data is retained.

Operational checks before acting on a flag

A useful flag should let an operator reconstruct the match quickly:

Before escalating a result, several checks help avoid false alarms:

The appropriate response depends on what the match represents. A known national campaign in two affiliates may require only a reporting update; an unexpected copy in a monetised library could need a rights check. Version clusters can also reveal whether a duplicate came from one master file or several separately edited copies. Keeping the evidence attached to the flag preserves an audit trail without forcing every match into the same workflow.

What to remember about cross-feed detection

ReCAP identifies reused commercials by turning moving images and sound into searchable evidence, then comparing that evidence across feeds with tolerance for normal delivery changes. Audio, visuals, recognised text, logos and sequence all contribute, while metadata and quality information help explain the result.

The essential point is that a credible cross-feed duplicate flag combines similarity with context. It shows where the content appeared, how closely it matched, which features supported the result and what remains uncertain. That combination separates machine evidence from human judgement and gives media teams a practical basis for spotting repeated advertising while respecting the legitimate local variations that shape the Australian market.