One of my current research is the use of Deep Convolutional Neural Networks DCNN for super scaled resolution. This provides a better way to use machine learning for visual content processing. Photoshop can also upscale images, using Bicubic Interpolation. The issue with that is it leaves the image less sharp and blurry and requires enhancing edges to retouch the image. There is more post work left to edit an image whereas using AI with DCNN, the image is not only superscaled to a higher resolution, but enhanced as well.
Practical application of this method will greatly improve photos uploaded on social media apps and as a filter for image editing programs. DCNN allows increasing the photo or image's size without losing quality. You lose less detail when downscaling, which is why when a video is shot in 8K resolution, it still appears clear and sharp in full 1080p HD resolution. In Photoshop when you upscale an image, you lose information in details which results in lossy resolution. Contrast also is reduced. You then have to retouch it more to get the quality and sharpness, but now it looks softer. With DCNN, you don't lose information, you actually create new details on the image to enhance it's quality. The neural net uses a dataset of values it has learned from other images and applies it to enhance and upscale the original image. Artifacts from lossy JPEG compression is removed along with digital noise and sharpens the edges. This can result in more sharper images that have been upscaled from low resolution photos. Those who use an app with a smartphone camera that is not quite high resolution can use this feature to get high quality images.
Photography credits Vicente Tabora
Further Reading:
https://en.wikipedia.org/wiki/Convolutional_neural_network
