Thanks to the technique ofupscaling, it is possible to improve the rendering of an old photograph in a realistic way. While knowing that artificial intelligence can “imagine” plausible details that are not exactly consistent with the original.
This will also interest you
[EN VIDÉO] Artificial intelligence that adjusts the angle at which the photo is taken Interactive point-based manipulation of the generative image collector. © DraGAN Project
Do you have an image that was scanned at low resolution or whose original was relatively blurry? Most of the time, such an image is difficult to use for professional use. Fortunately, artificial intelligence has come to interfere in this territory. There are many so-called applications upscaling (scaling) which improve the resolution of an image, using artificial intelligence. LetsEnhance, Cutout Pro, Upscale Pics…
The “upscaling” technique
Usually, image editing programs can improve the resolution of an image by “interpolation.” In this scenario, the program uses a mathematical method to estimate intermediate values between several points. For example, a missing pixel is evaluated from the 4 closest pixels, or even from the 16 closest pixels.
Interpolation can certainly help improve an image, but it cannot add realistic detail to it. In fact, if we enlarge said image, the effect of the interpolation may not be very graceful. This is where AI can do its job.
L’upscaling relies on deep learning, or deep learning in English, and strives to improve an image by offering to complete it in a realistic way. In other words, theupscaling “imagine” the missing details of the image.
To this end, an application of upscaling has usually been trained on hundreds of thousands or millions of high resolution images. Said images have been intentionally reduced in order to be converted to low resolution. The application then learned to reconstruct a high-resolution image from its low-resolution equivalent. From this training, the applications of upscaling can therefore start from a low resolution image and “predict” its high resolution equivalent.
A practical example
Here we have an example for which we took advantage Let’s Enhance. We started with a low resolution pixelated image.
We loaded this photograph into Let’s Enhance and asked for it to be improved – without changing the settings offered by the software. Here is the result, below.
Upscaling does not necessarily reproduce the original as is
The resolution of the photograph has been greatly improved and, in addition, Let’s Enhance has made the swimmer’s limbs much sharper. It should be noted that the artificial intelligence used here imagines realistic shapes for the retouched photograph but does not necessarily reproduce the original image. In the case of this photograph, the swimmer was wearing a red swimsuit with white polka dots. The application has imagined a realistic pattern, but which is therefore not exactly true. Here we have a detail that is good to take into account when improving the resolution of an old photograph in this way.
rewrite this content and keep HTML tags