ReCAP and Animated Lower-Thirds in Broadcast News

Animated lower-thirds in television news carry names, roles, locations and story labels while remaining on screen through wipes, slides, type-on effects and layered transitions. Treating them as ordinary subtitles produces unstable transcripts, duplicate identities and false matches. It also confuses editorial content with scores, tickers, captions and sponsor bugs that share the same pixels. A usable archive must understand the graphic as a timed object, not as a collection of isolated words.

The EU-funded Real-time Content Analysis and Processing (ReCAP) initiative is suited to this problem because it combines real-time analysis of broadcast-quality video with metadata extraction, quality monitoring, face and logo recognition, and duplicate-content detection. ReCAP’s approach to recognising and logging animated lower-thirds can connect what viewers saw with who or what was named, when it appeared, and whether the source footage remains trustworthy. A lower-third is both evidence of speech and a component of the visual package that can identify a programme, contributor or rights holder.

Why moving name straps defeat simple text capture

Conventional optical character recognition works best on stable, high-contrast text. During a lower-third animation, letters may appear in fragments, move behind a coloured panel, fade through transparency or disappear beneath another wipe. Reading every frame can split one person’s name into several entries; reading only the final frame can miss spelling changes and editorial corrections. Recognition therefore needs a stable reading moment chosen from the whole animation. For an archive, these errors become increasingly damaging because subtitling, search and rights workflows may reuse flawed text without knowing it came from a transition.

Spatial and temporal rules can narrow the search without depending on one fragile cue. ReCAP can treat the lower portion of the frame as a candidate zone, then combine text edges, bounding boxes, contrast, motion, opacity and continuity with the programme’s graphic layout. A sports score bug or closed caption may share some properties, but it differs in position, scale, update rhythm or persistence. Confidence comes from agreement across frames rather than a single OCR result. A rule-based template can help when a station package is consistent, although it should be configured for each channel rather than treated as universal.

Public web corpora also contain noisy material; a page about Telegram search optimisation may be indexed beside genuine broadcaster references, so source authority and capture context must be retained with any logo match. This matters when a network mark, sponsor logo or public figure appears briefly inside a kinetic graphic and the recognition system consults external visual evidence. The same principle applies to screenshots and clips harvested from social platforms, where the original publication context may be missing or fabricated.

A ReCAP pipeline for detection, recognition and timing

This approach supports the wider goals set out in the ReCAP project objectives, where automatic metadata extraction sits alongside content recognition and quality monitoring. The lower-third pipeline begins at ingest, where proxy frames, programme timecode, audio timing and source identifiers remain linked. Frames can be normalised for resolution and contrast, while the original stream or evidence frames are preserved for later review. Processing should be frame-accurate, but full-resolution recognition does not need to run on every pixel of every stream.

A tracker then classifies each graphic’s lifecycle: absent, entering, active, changing and exiting. OCR is sampled when the lettering is most stable, but the engine records earlier and later observations so a correction from “live” to “final” is not silently overwritten. A change detector can trigger full OCR after a wipe completes, while a lightweight background pass watches for later edits or superseding straps. The system also distinguishes a name strap from adjacent text, ticker crawls, maps and sponsor bugs, retaining separate confidence values for detection, reading and semantic assignment.

The output is an event record rather than a pile of OCR lines. It can contain the graphic’s start and end times, timecode range, reconstructed text, language, subject type, associated person or organisation, and links to source frames. Where consented face matching is available, it may support a speaker association, though the visible strap and editorial metadata remain stronger evidence than an automated facial guess. An editor can then approve or amend a proposed subject link without changing the immutable detection record.

Turning pixels into an auditable media record

Each detected animation should receive a persistent graphic identifier and a searchable normalised title, with raw readings kept beside the corrected value. This lets a producer retrieve “Jane Porter — Chief Meteorologist” without collecting every transitional fragment, while an editor can still inspect the original spelling and its evidence. Useful fields include programme, story, frame rate, duration, language, confidence, extraction method, review status and checksum. A searchable alias list can connect spelling variants without replacing the text actually broadcast.

Automated web matching needs equally strict provenance. A result calling itself an online casino assessment should never be accepted as evidence for a broadcaster or sponsor identity merely because its wording or favicon resembles the graphic. Captured pages, reference logos and source dates should remain attributable, allowing operators to reject irrelevant or manipulated matches without blocking the entire recognition task. Similarity scores should guide review rather than make the final identity claim.

Logging also supports quality control. Duplicate or near-identical straps can be grouped, while missing text, unstable timing and inconsistent names can be flagged for review. Human corrections should be versioned rather than pasted over machine output, with retention periods aligned to licence terms and privacy obligations. Access controls can separate raw facial evidence from searchable editorial metadata where the system supports that separation. This audit trail turns an experimental detector into an operational media record that can withstand legal, editorial and technical scrutiny.

Australian newsroom conditions shape the workflow

Australia’s broadcast market combines large Sydney and Melbourne production centres with free-to-air ABC and SBS services, commercial channels, Foxtel operations and regional stations. Each may use different graphics packages, playout systems and archive platforms. A shared recognition layer can describe them all while preserving source-specific templates, making search records more useful across fragmented teams and deadline-driven workflows. Cloud playout, shared edit systems and fast-turnaround social clips can also multiply copies before they reach the archive.

Language and naming conventions add another layer. SBS bulletins may carry names and roles in more than one language, while Australian coverage frequently features Aboriginal and Torres Strait Islander names, overseas visitors and unfamiliar place names. Preserving diacritics, transliteration and alternate spellings is essential; normalising a name should not erase cultural specificity or make two people appear identical. Australian English and local place naming need equally careful handling. An unfamiliar surname is not evidence of an error, so the catalogue needs confidence labels rather than silent corrections.

Scheduling must also respect local realities. Sydney and Melbourne observe daylight-saving changes, while Perth and other western centres run on a different clock, so every event can link editorial local time, UTC and frame-accurate timecode. That precision helps during breaking coverage, elections, severe weather and bushfire bulletins, when name straps change quickly and several producers may need to find the exact graphic within seconds. Emergency graphics also need high-contrast text and clear timing for viewers using captions or low-bandwidth streams.

Recommendations for reliable lower-third logging

A reliable deployment should make its evidence, uncertainty and operating rules visible from the outset. The following controls are especially valuable for teams moving between public, commercial and regional workflows.

Capability Frame-only OCR ReCAP-oriented temporal workflow Newsroom value
Animated text Captures temporary fragments Tracks states and reads stable frames Fewer duplicate names
Identity link Depends on one clear image Combines strap, context and approved matching Better subject retrieval
Timing Often limited to a frame or interval Preserves lifecycle and source timecode Faster archive searches
Quality Leaves errors embedded Flags uncertainty and review needs Stronger editorial trust
Provenance Little source context Retains evidence and version history Auditable reuse

These controls should be evaluated together because stronger OCR alone will not solve animation, timing or provenance. A useful pilot measures whether editors can retrieve the right person, title and graphic faster, whether transcripts remain stable across replays, and whether low-confidence cases reach a person before publication rather than after an error. It should also establish whether automated records remain consistent when the same bulletin is sourced from transmission, post-production and an online replay.

The next concrete step is to nominate one Australian half-hour bulletin, label its animated lower-thirds as a test set, and run a ReCAP pilot against the original video and graphics metadata to measure recognition accuracy, timing consistency and editorial search time.