Using AI-Powered Photo Enhancement for Open Source Intelligence

If you are active on social media, there is a good chance you have seen this viral video of an 1896 snowball fight. A clip originally shot by the French brothers Lumiere, the footage has been scaled up using AI, colorized and now provides for a pleasant 50 second relief from your daily routine of compulsive doomscrolling.

 

 

As an OSINT aficionado, what got me much more excited than the frolicking fin de siècle Frenchmen was the fact that the image had been scaled up using commerically available AI image enhancement software. Conventional wisdom holds that you cannot really improve the quality of an image. The image holds a certain amount of data and no amount of fancy technology will ever be able to pull additional data out of thin air. But as anyone who has ever processed or analyzed a satellite image knows, the question often is less how much data an image holds and more how much of the data you are able to retrieve. Could AI-powered software help squeeze even more information from the tired old jpeg and potentially do so in a visually appealing fashion?

In order to find out, I tried out Gigapixel AI, one of the off-the-shelf software options available from the company Topaz Labs. Topaz offers several tools for AI-powered image enhancement including image sharpening, video enhancement and image upscaling which was used here. And in case you wondered, much to my chagrin, I am not commercially associated with Topaz Labs in any way.

So let's try a classic example of OSINT nerds suffering from a lack of image quality. This picture from March 2017 shows Kim Jong Un right next to a Hwasong 14 engine. The original image is rather small and it would be great to have a higher resolution image showing more of the engine's details. A double fold increase in size using Gigapixel AI results in a much clearer image of the engine. It's no magic, but still quite an improvement over conventional upscaling.

But then again, this is a hard case. The engine section on the picture is pretty dark and the original image is really quite tiny. To see what happens if you apply Gigapixel to a higher image quality baseline, I ran the program on one of the high-res pictures of North Korea's giant new ICBM first posted by Ankit Panda.

The result is pretty amazing. While some rivets are exaggerated and there are some minor artifacts, the scaled up version retains main shapes and edges without major quality loss or blur. Details that one would have guessed correctly after staring at the original pic for several minutes are now fully visible in a clear and crisp image (for the full original image click here and for the scaled-up version here).

Gigapixel's ability to scale up images while retaining shapes and revealing hard to discern details is great for parade pictures but there is one area where these capabilities come in even more handy – satellite imagery.

Applying Gigapixel to standard Google Earth imagery can be a bit fidgety. Since the software needs to recognize edges, a smaller but sharper image often yields better results than a larger picture that is slightly blurry. The below images show a Syrian military production site near Masyaf  (35.154° 36.344°) both in the original version taken directly from Google Earth and scaled up using Gigapixel AI.

Once scaled up, the image does begin to look a little funny. The overall impression gets closer to an acrylic painting than a satellite image and some of the details start to look a little weird. However, the basic shape of the building is not only retained but its edges actually get sharper during the upscaling process. With measurements being quite sensitive to both resolution and sharpness this already seems like a pretty useful capability to have.

Since Gigapixel allows you to scale up an image with reduced loss of quality, another question arises. Can you scale up a satellite picture and then scale it down again to create an image of the same size as the original but of slightly higher quality? Well, turns out you can. If your start from an already expertly processed high-end image, the effect might be quite subtle. But if your baseline is just regular Google Earth imagery, the results can be stunning. Below is an original Google Earth picture of Masyaf right next to the version scaled up using Gigapixel AI and scaled down again using Photoshop.

And here is an image showing a section of North Korea's sprawling Hamhung chemical complex (39.848° 127.586°) processed using the same method.

In case you wondered how this form of processing compares to regular sharpening, here is a comparison using Sudan's mystery facility (16.253° 32.642°). For sure, using Gigapixel AI is a bit more work than just pressing the Photoshop sharpening button but if you, like me, are obsessed with the crispiness of your sat imagery, there is a clear winner. 

In conclusion, Topaz' Gigapixel AI seems like a pretty useful new tool for use in OSINT in general and the processing of satellite imagery in particular. Unfortunately, due to limited computing power at my disposal, I haven't tried out Topaz' Video Enhance AI yet. If any of you is in possession of more advanced tech and wants to give it a try I'd love to hear about your results! As always, corrections, constructive criticism and incoherent yet hilarious rants are welcome.