From small to not so pixel-perfect large

Everyone knows this problem: a friend sends a low-resolution image of last weekend´s hike to your smartphone, but when you save the picture of the beautiful bird and later add it to a digital photo album, the image shows checkerboard artifacts. The resolution is just too low. In times of need software is utilized that promises to upsample small sized images, but with poor results: Those holiday pictures look blurry and lack high definition.

Photo: EnhanceNet-PAT is capable of upsampling a low-resolution image (left) to a high definition version (middle). The result is indistinguishable from the original image (right).

Technology to create a large size from a low-resolution image is known as the single image super-resolution or SISR technology. SISR has been studied for decades, but with limited results. The software adds extra pixels and fills them with the average “look” of all the surrounding pixels. The result is blurriness. Researchers at the Max Planck Institute of Intelligent Systems propose a new approach to give images a realistic texture when magnified from small to large – through the help of Machine Learning. Artificial Intelligence is at play, where the algorithm for upsampling the image learns from experience in sharpening its look.

The learning process is much like that of a human: practice makes the master. “The algorithm is given the task of upsampling millions of low resolution images to a high resolution version and is then shown the original, the “this-is-how-it-should-be”-image. Notice the difference? OK, then learn from your mistake”, explains Mehdi S.M. Sajjadi, who together with Dr. Michael Hirsch and Prof. Dr. Bernhard Schölkopf, Director of the Empirical Inference Department at the Max Planck Institute for Intelligent Systems in Tübingen, developed the EnhanceNet-PAT technology. Once EnhanceNet-PAT is trained, it no longer needs original photos.

When EnhanceNet-PAT is put to work, according to the researchers, the technology is more efficient than any other SISR technology currently on the market. The difference lies in the pretense of wanting to be pixel-perfect. In contrast to existing algorithms, EnhanceNet-PAT gives up on pixel-perfect reconstruction, but rather aims for faithful texture synthesis. By being capable of detecting and generating patterns in a low resolution image and of applying these patterns in the upsampling process, EnhanceNet-PAT thinks how the bird´s feathers should look like and adds extra pixels to the low-resolution image accordingly. You could say the technology created its own reality. For most viewers, the result is very much like the original photo. The picture of the bird is good to adorn the photo album.