Abstract:

The aim of this work is to develop an interactive tool where you can see the effect of different parameters related to the artificial intelligence (deep learning) Style Transfer technique in a simple and intuitive way. The idea is to allow users to play with the different parameters of the Style Transfer technique and see how they influence the result of its application. This allows users to visualize in a simple way how these parameters relate to each other, and how they affect the transfer of the style.

Artificial intelligence (AI) is one of the fastest-growing fields in recent years. AI is the science that seeks to mimic human behavior in some way. Within AI we have the field of machine learning (ML). ML involves a set of techniques that allow machines to learn from data, obtaining AI applications. The results obtained depend heavily on the data, and it is necessary in many cases to perform preprocessing to obtain the intrinsic features of the data. This is where deep learning (DL) arises, where the machine not only learns to use the data but also to extract its features from the data beforehand. It is called "deep" because it uses a technique called neural networks, with many layers, that is, a lot of depth.

One of the best-known examples of DL is image classification. A classic use case is to detect whether there is a cat or a dog in a picture. Another example is the generation of synthetic data or images, such as deep fakes, where you generate images that look real but are not. There is also Style Transfer in the category of image generation.

Style Transfer is a deep learning technique where the style of an image is applied to another image from which the content is taken. That is, certain features of an image, such as textures or color, are transferred to another image from which particular information such as objects or the scene is retained. This allows, for example, to obtain an image of the Mona Lisa as if it had been painted by Picasso. Three factors have a great influence on this and are closely related to each other. These are learning rate, style weight, and content weight.

The style weight, in the case of the Mona Lisa painted by Picasso, would indicate the freedom that Picasso has to apply his famous style. On the other hand, the content weight would be related to how much it should still resemble the original Mona Lisa in terms of content, that is, how important it is to keep all the objects and general structure of the painting. Finally, the learning rate has an effect on the learning speed, which in practice means a different result in terms of the smoothness of the generated image.

So, the main idea of the web dashboard tool we present in this work is to visualize the effect of these parameters in a practice way. The dashboard tool permits to load a real cine cardiac magnetic resonance image as an example. It also has a synthetic blurred image where the singular style of the real image is applied. The user can manipulate three sliders with which the values of the parameters can be specified. Tuning the mentioned parameters, the user can visualize the resulting image and compare it with the original and the blurred image.

This tool allows users to understand the effect of the parameters related to the otherwise complex Style Transfer technique. It makes it an intuitive and attractive tool to help newcomers to this field, and it shows a potential application in medical imaging research.

Keywords:

Educational tool, deep learning, image analysis, Style Transfer, dashboard, image processing, medical imaging, artificial intelligence.