Although YouTube has implemented its own measures to prevent brands’ ads from appearing on or near objectionable content, marketers’ battle for brand safety on that popular video platform is a long ways from being over.

Now there is another tool in that struggle. Los Angeles-based social entertainment company Fullscreen is undertaking a new direction with the launch of its own brand safety tool for YouTube videos.

How the tool came about. The company’s regular job is matching brands with influencers, providing production crews to the influencers if needed, and then placing ads developed from the resulting collaboration as pre-rolls on YouTube. Additionally, it operates as a media buying agency to place others’ ads on that social platform and others.

The need to create a new brand safety tool for YouTube, senior director of media operations Mark Williams said in an interview, begin in March of last year following the “Adpocalypse,” which saw the abrupt exits of major brands from YouTube over ad placement near objectionable videos.

The wave of departures affected Fullscreen’s business, he said, as brands “started to pause,” and as his company “saw our ads running on places we didn’t want them to.”

How it works. Called Shield, the new tool detects scene changes in videos, grabs a screenshot from each scene, and then conducts image analysis on the screenshots. The company says the platform can recognize up to 4,600 objects, a number that is growing via machine learning training.

Shield also performs audio-to-text transcription of the soundtrack, and captures the metadata associated with each clip. Shield is able to translate more than five dozen languages.

These textual files — descriptions of objects, audio-to-text, and metadata that includes comments and audience ratings — are then analyzed against a blacklist of words representing brand safety minefields, covering terrorism, extremism and obscenity. Shield employs Google’s speech-to-text processing and its image processing, while a Fullscreen proprietary engine detects the scenes, pulls the shots and does the textual analysis.

Currently, the evolving blacklist contains about 12,000 keywords. Williams said a YouTube video of a few minutes in length can be processed in about 30 seconds, while a scan of videos needed for a campaign — as many as 25,000 clips — might take up to 48 hours.

How it categorizes content. Since the ads delivered by Fullscreen for its clients are not handled by programmatic platforms, the brand safety processing doesn’t need to operate in milliseconds. The company is planning to soon utilize the tool for Instagram videos.

Videos are classified in four conventional TV categories: G, PG, 14 and MA (Mature Adult). If the machine can’t categorize a video, such as when the images are blurred or the sound indistinct, there is a manual review by a human, and that instance is used to help train the engine. Williams said the accuracy data won’t go beyond anecdotal evidence until a just-completed large test with a telco client is completed.

Why this matters to marketers. As with other marketers, Fullscreen became frustrated with the options available for brand safety scanning on YouTube.

Williams said his company’s search for a tool to remedy the problem turned up ones that mostly relied on text or logos. Other tools offer scene recognition, such as Clarifai, and Williams points to competitors like Studio 71 which offer image recognition, but he contends Shield provides a more comprehensive analysis of video image, soundtrack and metadata than other approaches, and is more transparent about the reasons for its scoring. Using its growing database of YouTube content, Williams said his company is also planning on developing a YouTube search and discovery tool.

Until YouTube has a platform-wide, fairly foolproof way to guarantee brand safety for advertisers, solutions like Fullscreen’s will continue to appear. Their appearances will likely flourish, as other agencies, brands and vendors take matters into their own hands, in part by piecing together the software-as-a-service capabilities from Google and others for image recognition, speech-to-text and related services.

This story first appeared on MarTech Today. For more on marketing technology, click here.