What makes the image memorable? Ask the computer. From "Mona Lisa" to "Girls with Pearl Earrings," some images linger in my mind for a long time after other photos disappear. Ask an artist why you may hear some universally accepted principles for making great art. Now, there is a more straightforward learning method: an artificial intelligence model is required to draw an example.

Artificial intelligence models show amazing details.
Artificial intelligence models show amazing details.


A new study uses machine learning to generate images, ranging from great cheeseburgers to memorable coffee, detailing why portraits or scenes stand out. The most famous pictures of individual objects in the study have vivid colors, simple backgrounds, and objects that are prominent in the film. The results were presented at this week's International Conference on Computer Vision.

Co-authors of the study, Phillip Isola, Bonnie and Marty (1964) Tennbaum CD, Assistant Professor of Electrical Engineering and Computer Science at the Massachusetts Institute of Technology, said: "A picture is worth a thousand words. ". "There are a lot of articles about memory, but this method allows us to see how memory looks visually." It gives us a visual definition that is hard to express in words."

This work is based on the early model MemNet, which scores the memory of the image and highlights the image features that influence its decision making. MemNet's predictions are based on the results of online research in which 60,000 images are displayed to human subjects and ranked according to how easy they are to be remembered.

The model analyzed in the current study uses a machine learning technique called Generating Against Networks or GAN to visualize a single image because its image is huge from mech to unforgettable distance. Analyze allows viewers to visually see that the fuzzy panda lost in the bamboo becomes a skeleton-based panda, with dark eyes, ears, and claws in sharp contrast to the white cup.

The image segmentation GAN has three modules. The movement-based evaluator can rotate the memory knob on the target image and calculate how to achieve the desired effect. The converter executes its instructions, and the generator outputs the final image.

Gradient has a dramatic feel for time-lapse images. The cheeseburger, which was transferred to the end of the memory level, looked fatter and brighter than the previous version and was "delicious," as the author pointed out. A ladybug seems more intelligent and more purposeful. Unexpectedly, the pepper on the vine changed from green to a red chameleon.

The researchers also studied which features have the most significant impact on memory. In an online experiment, images with different minds were displayed to human subjects, and any duplication was marked. It turns out that the most viscous replicas bring the issue closer, making the animals or objects in the frame appear larger. The next most crucial factor is the brightness so that the subject is in the center of the picture and is square or circular.

Research co-author, Lore Goetschalckx, a visiting graduate student at Katholieke Universiteit Leuven in Belgium, said: "The human brain has evolved to focus on these functions, which is what GAN is pursuing."

The researchers also reconfigured GANanalyze to produce images with different aesthetic and emotional appeal. They found that from an artistic and psychological point of view, the higher-rated models were brighter, more colorful, and had a shallower depth of field, blurring the background, just like the most memorable pictures. However, the most beautiful images are not always significant.

Researchers say that GANalyze has many potential applications. It can be used to detect and even handle memory loss through object enhancement in augmented reality systems.

Instead of using drugs to enhance memory, you can improve the world with augmented reality devices and make things that are easy to misplace.

analyze can also be used to create great graphics to help readers retain information. "This can revolutionize education," Oliva said. Ultimately, GAN has begun to generate synthetic, realistic world images to help train automated systems to identify locations and objects that are unlikely to be encountered in real life.

Post a Comment

Previous Post Next Post