Neural Network Super Resolution getting good, fast

Gepostet vor 11 Monaten, 6 Tagen in #Design #Science #Tech #AI #AlgoCulture

Share: Twitter Facebook Mail

enhance

Neulich bloggte ich über die K.I.-Version der Zoom & Enhance-Meme und meinte noch, das seien ja nur 64x64 Pixel und die Ergebnisse wären ja noch eher so mittel. Scratch that. Ein neues Paper stellt eine neue Methode (SRGAN) zum neural-network-gestützten hochrechnen von Bildern vor, das dramatisch bessere Ergebnisse erzielt, in dem es Informationsverluste in den Texturen durch gelernte Algo-Pattern ausgleicht. Zoom & Enhance is coming!

enhance2

Despite the breakthroughs in accuracy and speed of single image super-resolution using faster and deeper convolutional neural networks, one central problem remains largely unsolved: how do we recover the finer texture details
when we super-resolve at large upscaling factors? During image downsampling information is lost, making superresolution a highly ill-posed inverse problem with a large set of possible solutions. The behavior of optimizationbased super-resolution methods is therefore principally driven by the choice of objective function. Recent work has largely focussed on minimizing the mean squared reconstruction error (MSE). The resulting estimates have high peak signal-to-noise-ratio (PSNR), but they are often overly smoothed, lack high-frequency detail, making them perceptually unsatisfying.

In this paper, we present superresolution generative adversarial network (SRGAN). To our knowledge, it is the first framework capable of recovering photo-realistic natural images from 4× downsampling. To achieve this, we propose a perceptual loss function which consists of an adversarial loss and a content loss. The adversarial loss pushes our solution to the natural image manifold using a discriminator network that is trained to differentiate between the super-resolved images and original photo-realistic images. In addition, we use a content loss function motivated by perceptual similarity instead of similarity in pixel space. Trained on 350K images using the perceptual loss function, our deep residual network was able to recover photo-realistic textures from heavily downsampled images on public benchmarks.

Algorithmic Image-Watermark Remover

Google hat ein neues Paper über die automatische Entfernung von Wasserzeichen in Pics, netterweise nennen sie das Paper „On the…

10 PRINT CHR$ (205.5 + RND (1)); on a Commodore Pet

10 PRINT CHR$ (205.5 + RND (1)); 20 GOTO 10 ist sowas wie ein Miniatur-Programm in Basic für Generative Graphics,…

DeepMind releases StarCraft AI

Google und Blizzard haben ihre StarCraft AI veröffentlicht: Testing our agents in games that are not specifically designed for AI…

Self-Driving Cars hacked with Love and Hate

Wissenschaftler der Uni Washington haben die Computer-Vision-Systeme von selbstfahrenden Autos gehackt – indem sie Love und Hate in colorierten Pixelfonts auf…

Cellular Automata Cube

Cubes.io: Conways Game of Life als 3D-Spielzeug mit Cubes und Spheres und Schnickschnack als Evolution-Nullpunkt, von wo aus die ganzen…

Neural Network-Faces synched to Music

„My first attempt to map a song made by @kamptweets onto GAN generated proto-faces.“ Bohemian Rhapsody next. The Three Nightingans.…

AI-Animations with human Sounds

Google vor ein paar Tagen so: „Yay, wir haben hier 'ne neue AI-based Animation-Tech, hooray!“ (Paper) Hayayo Miyazaki über AI-based…

Visual AI-Spaces Auto-Pilot

Ich habe schon ein paar mal über Mario Klingemanns Arbeiten hier gebloggt, derzeit jagt er Neural Networks durch Feedback-Loops und…

Synthesizing Obama from Audio

Im Mai bloggte ich über ein damals noch nicht veröffentlichtes Paper zur SigGraph2017, in dem sie eine Methode für generative…

Generative Pearls

Cool fractal and generative art by Julien Leonard. I dig his explanation from his about-page: „I create algorithms that connect…

Moarph

Mario Klingemann does some weird shit again with CycleGAN Feedback Loops (Neural Networks feeding their results back to each other).…