Aug 31, 2021 11:59:00

Google improves the performance of 'AI model that converts low resolution images of Gabigabi to high resolution images', to a level that humans can not distinguish

Google's AI research team

Google AI 'is, dare add noise to the low-resolution image is processed until the' pure noise ', to generate a high-resolution image from which diffusion model (diffusion model) technique called' Announced a new approach to improving. 'Technology for generating high-resolution images from low-resolution images with poor image quality' is expected to be used for a wide range of purposes, from restoring old photographs to improving medical images, and is one of the tasks expected to be useful for machine learning. ..

Google AI Blog: High Fidelity Image Generation Using Diffusion Models
https://ai.googleblog.com/2021/07/high-fidelity-image-generation-using.html

Enhance! Google researchers detail new method for upscaling low-resolution images with impressive results: Digital Photography Review
https://www.dpreview.com/news/0501469519/google-researchers-detail-new-method-upscaling-low-resolution-images-with-impressive-results

Google's New AI Photo Upscaling Tech is Jaw-Dropping | PetaPixel
https://petapixel.com/2021/08/30/googles-new-ai-photo-upscaling-tech-is-jaw-dropping/

In general, the task of restoring high-resolution images from low-resolution images uses generative models such as hostile generation networks (GANs) , variational auto-encoders ( VAEs), and autoregressive models. However, GANs have some drawbacks, such as mode collapse in which many of the generated images are duplicated, and autoregressive models have problems such as slow synthesis speed. And that.

On the other hand, the generative model called 'diffusion model' announced by Google AI in 2015 seems to have been reviewed in recent years due to its stability in training and the high quality of generated images and sounds. And newly, Google AI will improve the image composition quality of diffusion models by using two new diffusion model approaches, 'Super-Resolution via Repeated Refinements (SR3) ' and ' Cascaded Diffusion Models (CDM)'. It states that it was successful.

First, SR3 gradually adds Gaussian noise to the low-resolution image and damages it until it becomes a 'pure noise image'. After that, the trained neural network reverses the image corruption process to remove noise and generate a high-resolution image that exceeds the original resolution.

The left is a low-resolution image of 64 x 64 pixels, which is the input data, and the right is a 'pure noise image' by adding Gaussian noise to the low-resolution image.

Gradually remove noise from the 'pure noise image' ...

Finally, a much finer facial photo than the original image was generated.

It seems that it works in the same way for female images.

This is the result of the research team actually showing a human subject 'the original image' and 'an image generated from a low-resolution image by various methods' and having them determine which is the original image. The closer the subject's error rate (confusion rate) is to 50%, the more difficult it is to tell whether the AI-generated image or the original image is genuine.

Compared with methods such as FSRGAN (Face Super-Resolution Generative Adversarial Network) , PULSE , and Regression (self-regression generation model), SR3 is 'when a 16x16 pixel image is made 128x128 pixels' (top) and ' The confusion rate is high in both cases of 'when an image of 64 x 64 pixels is changed to 256 x 256 pixels' (below). In particular, the confusion rate when the 16 x 16 pixel image was changed to 128 x 128 pixels was 47.4%, and it seems that the subject could hardly distinguish the image generated by the AI model.

Google AI is also announcing CDM, a class-conditional (labeled) diffusion model trained on ImageNet for large image recognition datasets. ImageNet contains a diverse set of data, which can result in images that are far from the original image, but CDM gradually upscales the generated model with label information at multiple spatial resolutions. By doing so, it is possible to generate high-quality images.