#Google reverse image lookup archive
#Google reverse image lookup tv

A clip purporting to be a CNN segment from last week could be instantly scanned against the real CNN footage from that day to confirm or refute the clip’s authenticity.

#Google reverse image lookup archive

Why might the ability to connect the broadcast and digital realms be useful?Īrchives like the Internet Archive’s Television News Archive could serve as canonical reference archives to verify unknown footage. In contrast, connecting television and online imagery represents a slightly different need in terms of the strictness of matching, suggesting a few additional filtering steps could help bridge this difference. Absolute exact matching is not its primary purpose. In short, Google's Vision AI API is designed for general purpose image similarity in which it is desired to use fuzzy matching to find different images that still look highly similar to the input image. Given that the Vision AI API also supports OCR text extraction, these additional details could be fairly readily incorporated into a hybrid visual-textual matching filter that would permit far more strict matches. This suggests that the subtler differences that similarity algorithms are designed to look past, such as textual overlays and the specific time and date displayed in the corner of the frame, are extremely important when it comes to distinguishing among the vast archives of extremely similar imagery that define television news coverage. This is made worse by the fact that the studios they present from are refreshed fairly rarely, meaning there are few visual cues to distinguish the passing of time. In fact, search on Google Images for the name of any major anchor and the resulting collage of imagery makes clear how similar their outfits typically are from day to day. Even as a human, a screen capture of Wolf Blitzer from a month ago would likely be indistinguishable from a capture taken this morning. Television represents a worst-case scenario for visual similarity matching in that the same newsreaders sit in the same studios year after year with the same clothes and same facial expressions performing the same task. There are also many matches of prominent newsreaders from the sample week matching against clips of them from previous years sitting in the same studio wearing similar outfits. More complex, however, is the example of footage of Trump in the Oval Office in April 2019 matching footage of him in the Oval Office from December 2018. These matches could be further tightened by applying a second filtering process that retains only precisely exact matches. In other cases, imagery of President Trump in the Oval Office from this past December on CNN was identified as strongly similar to imagery of him in the Oval Office during the sample week.Įven though these results were based on the Vision AI API’s narrow list of likely exact matches, rather than its much broader “similar images” list, the API still permits a fair amount of difference in order to maximize its ability to identify images that have been slightly altered. Some tweets included clips of broadcasts from the sample week or just before it which were still airing, like this Fox News clip. Interestingly, the API identified 229,344 tweets that contained images it believed were strongly similar to broadcast frames. The low overlap appears due to the fact that Google seems to have searched only a small number of search results pages, rather than indexing the Archive's thumbnail collection.

#Google reverse image lookup tv

The Vision AI API also identified 23,660 images overlapping with the Internet Archive’s TV News Archive that was used as the source for the experiment.