Skip to content

This AI is capable of removing unwanted objects from your photos, who said Photoshop?

26 mayo, 2021

Ruining a perfect photo is as simple as having a car appear at the most inopportune moment or a person passing by while you photograph a landscape. Removing them is relatively easy using Photoshop and the clone plug, but why use Photoshop when a neural network can do it for you?

DeepAngel is an Artificial Intelligence developed at the MIT Media Lab that, in short, is capable of remove unwanted objects from an image thanks to a neural network architecture based on Mask R-CNN and Deep Fill. Through its website, any user can upload a photo and delete almost any object they want, be it a person, a cat, a car or even a spoon.

ZAO, the Chinese MOBILE APP that through DEEPFAKE turns you into DICAPRIO in SECONDS

We tested DeepAngel, is it working well?

You can test the AI ​​for yourself as many times as you want. You simply have to access their website, click on “Erase with IA” and select which object you want to delete. Done this, you can upload an image from your computer or import it from Instagram. After a few seconds, the AI ​​will return the edited image (with better or worse result).

We have tested AI with different images and, although it is true that in some cases the results are far from perfect, they are still curious. Editing improves when objects are far away or do not occupy a large percentage of the image. Here are some examples below. On the left, the one edited by the AI; on the right, the original.

Plane Saving the reflection in the glass of the building, it is almost impossible to differentiate the original photo from the one edited by the AI.
Dog Although you can see a small distortion in the area where the dog was, if we did not know the original image it would be difficult to discern whether it has been edited or not.
People This time, you can see the edit perfectly. However, the result is really decent. Possibly, if the fence did not have the decoration that it does, the final image would have been much better.
Person In A Car For this image, we have asked the AI ​​to remove the boy playing the guitar. The result can be improved, but it is curious how it has managed to detect the person and leave the rest of the scene unchanged.
Spoon Ok We couldn’t resist trying the “Spoon ” category. The result is very improvable, although if it was placed on a completely white background, it would possibly have disappeared completely.

Although the result is interesting, the AI ​​commits some blunders in some images. As we noted earlier, if the object is too large or occupies a significant percentage of the scene, mainly because it does not have too much context to analyze. Here are two of the worst editions we’ve ever gotten.

Bike What better way to remove the bike than to put a huge gray square on top?
Cat As in the previous image, as the cat occupies most of the scene, the AI ​​is not able to analyze the context and returns a very improvable edition.

The creators of DeepAngel plan to present their project at the South by Southwest (SXSW) 2019 in order to warn of the application of AI tools in the manipulation of images, videos and photos, its use in the media and its corresponding impact on society. As it is said in the Book of Ecclesiastes, “once you are able to see what is not seen, your next task will be to test whether what you have seen is real or imaginary.”

There are tools capable of placing the face of one person on another to create false files (known as Deep Fakes), in the same way that there is software capable of manipulating videos and imitating the voices of politicians to make them say things they haven’t said. A bad use of this technology could not come to fruition.

In Genbeta | Magnifying tiny images is no longer a CSI trick thanks to neural networks