Yvan Cohen
Thu Feb 22 2024
AI and Digital Asset Management: Beyond the Hype
Photo by cottonbro studio
AI has erupted into our lives. Just a year ago it felt like the stuff of science fiction, today it is everywhere. AI can write a poem, plan your day, your business strategy and even your holiday. AI will push the frontiers of science and, yes, it might end up taking your job.
If AI is likely to shape our future, how might it change the way we manage digital archives?
At LightRocket, we have always understood digital asset management to be a movable feast. There is no finite solution. Solving problems is a learning process. Removing one obstacle sometimes reveals another. Technology evolves, business practices change, goals shift. Our challenge (passion even), is to respond and innovate.
Technology is like fashion. And right now, AI is in vogue. Everybody has an opinion on it, we are all trying to understand it and everybody wants a piece of it. If you want to turn heads in the tech world, sprinkle your website with references to AI.
Digital Asset Management and AI
When it comes to media management, the most obvious application for AI is in picture tagging and facial recognition. If you believe the hype, AI will automate your tagging and make finding the right picture easy. The reality is more nuanced.
The core technology behind AI tagging has been around almost a decade. Companies like Facebook were among the first deploy AI at scale. Facial recognition meant that Facebook's users could easily tag and find pictures of friends and family.
In essence, the process is simple. Technology is deployed to analyse images. Learning as it goes, the system gradually gets better at identifying objects, expressions and faces. The more pictures the machine sees, the more accurate it becomes.
Before I get to the bit where I challenge the hype, let me first say that AI is a pretty amazing tool, and for a certain level of tagging it works quite well.
Photo by cottonbro studio
AI Tagging and Facial Recognition
At LightRocket, we ran a series of tests using Amazon's AI technology set to a 'confidence' level of 80%. The results were impressive, if not completely convincing. Sometimes AI made some amusing gaffes, like mistaking a piece of cheese on a plate for a fried egg.
Facial recognition is perhaps the most popular and effective application for AI tagging. The structure of our faces constitutes a unique signature that AI learning tools can quickly identify. All we need to do is put a name to the face once, and every subsequent time your system sees that face, it will add the corresponding name.
If you're looking for a quick fix when it comes to tagging then, yes, AI can correctly identify elements in an image and can instantly add a bunch of corresponding tags. Which is probably (and only 'probably') better than having no tags at all. Facial recognition is particularly powerful, if putting names to faces is what you need - and questions of privacy aren't a major concern.
Now for the caveats.
All things being equal, AI tagging sounds like a perfect panacea for an otherwise time-consuming and labour-intensive task. The problem, however, is that AI only sees objects. It doesn't understand context, relevance or meaning.
The Short-Falls of AI in Digital Asset Management
To illustrate my point, if AI sees a person lying on the ground, how can it know if that person is resting, asleep or dead? It sees only the object.
If AI sees two people holding hands, AI cannot know if those people are lovers, friends or family. If it sees a picture of a woman cradling a child, it can't know if the mother is an HIV patient cradling her newborn child. Only the creator of the image can provide that information, which is why captions and comprehensive metadata (the who, what, why, when and where of a picture) are so important - and why our LightRocket interface places such emphasis on prompting users to complete that info.
Unable to understand context or meaning, all objects are of equal importance to an AI tagging machine. If you have a picture of a vintage car parked in front of a tree beside a post box, the tree and the post box are almost certainly not relevant to the meaning of the image (which is of a vintage car). But how will AI know if it should include, or not, the post box and tree in the tags? It doesn't.
The point with tagging, which people often miss, is that it's not about having lots of keywords to try and catch as many searchers as possible. It's about tagging your file with a few very relevant, carefully selected terms, so a searcher finds exactly the files they are looking for. It's about quality, not quantity. So, while AI tagging may be pain-free and fast, it probably isn't going to deliver the end result you are looking for: an accurate search.
As to facial recognition, there is no doubt that it can be useful and hugely efficient to have your system recognize repeat instances of a face. But there are caveats here too.
Photo by Tara Winstead
AI Tagging Isn't a Flawless System
Many organisations aren't prepared to take the responsibility for storing what are effectively facial imprints of individuals without their express permission - an issue that is especially pertinent in our privacy-conscious world.
For commercial or enterprise users, there's a second problem with facial recognition. While identifying somebody may be helpful, the significance of an individual within an organization is often linked to their role.
It may be great, for example, to be able to find all the pictures you have of 'Mr Brown' but what is Mr Brown's role? Herein lies the rub. If you add a role to an identity, it will be assigned to all instances of that individual. But what happens if the role of that person changes? You are left with files that are tagged accurately from a name perspective (most people don't change their names) but which may quickly become inaccurate as roles change.
There are of course other ways in which AI enhance search. One approach involves adding AI picture analysis as a search option. Using this approach, AI doesn't add tags to your files but delivers up search results based on search terms that correspond to content AI has identified within an image. So, if you want a picture of people smiling, an AI search could scan your collection to find those images without the need to add tags. It's another attractive-seeming quick fix, but it still doesn't resolve the issue of relevance and meaning, which is a far more valuable way of tagging images.
The Bottom Line
Though metadata management may not be the most joyous of tasks, there is some comfort in knowing that, for the moment at least, human analysis and choices are still the best starting point for effective tagging and accurate image retrieval. AI can certainly help and it certainly has its place in the world of digital asset management, but it hasn't taken the place of humans…yet.
Contact me, Yvan Cohen, at info@lightrocket.com if you'd like to find out more about how LightRocket Enterprise manages security.
Written by Yvan Cohen | Yvan has been a photojournalist for over 30 years. He's a co-founder of LightRocket and continues to shoot photo and video projects around South East Asia.
