Artificial Intelligence Working on Optical Illusions

An optical illusion is a type of image that generates a particular visual perception, making things appear that do not actually exist. In other words, the receiver is deceived by the visual system and believes in the optical effect created. Illusions are perceived naturally or under the influence of artificial substances (drugs, hallucinogens, etc.) in the existing environment. They are also caused by special geometric constructions, such as impossible objects, anaglyphs, contrasting compositions in op art, etc.

An impossible object

In recent years, the field of artificial vision has been studying this optical phenomenon. In 2018, researchers at Brown University trained a machine to see optical illusions as humans do. The technology presented in this article focuses on the illusions generated by colour contrasts.

The Café Wall Illusion

A study was recently conducted by a team from Flinders University in Australia on the café wall illusion. Neuropsychologist Richard Gregory was the first scientist to study this illusion, in reference to a tile wall cladding on the outside of a Bristol coffee house. This illusion makes parallel lines appear like curves.

The café wall illusion

The study centres on the design of a computer vision model that detects optical illusions and dissects their physiological origins. In addition to the café wall illusion, the researchers tackled other similar and more complex representations.

An Approach Based on Human Vision

Artificial vision technologies capable of detecting images like humans do are based on cognitive and biological studies of natural perception. The team used the physiological approach to mimic perceptual illusions. The researchers pointed out that even though scientists have made great strides in recent years in understanding the biological mechanisms of perceptual illusions, some aspects of stimuli processing remain controversial.

Before reaching the brain, images perceived by humans pass through the retina. Retinal ganglion cells are neurons that receive visual information and route it to the visual cortex. The size of each ganglion cell differs depending on its distance from the fovea. The fovea is the area of the retina where vision is the most accurate and detailed and from which most visual information is transmitted to the brain.

In order to mimic human vision, artificial vision models perform simultaneous image samplings on several scales, according to the principle of alternating receptor layers of visual stimuli in the retina.

Model Design Based on Lateral Inhibition

When faced with a visual illusion, our brain receives incorrect information. As it tries to interpret the information, it creates a different representation of the actual image. In the case of the café wall illusion, lateral inhibition sends the wrong signals to the visual cortex.

Lateral inhibition is “the selective suppression of certain nerve signals.” Its goal is to improve contrast perception by ensuring that “only the signal of well-lit photoreceptors is transmitted to the ganglion cells.” Photoreceptors are photosensitive neurons located in the outer layer of the retina.

In this study, the researchers evaluated a computer-based filter model designed to learn lateral inhibition and retinal ganglion cell responses to different geometrical illusions. To do this, they used difference of Gaussian, an image processing algorithm that enhances contours and details.

According to the researchers, the model is complex at this point and more research needs to be done to improve it. They also emphasized the importance of optical illusions in understanding and simulating natural vision.

The study entitled “Informing Computer Vision with Optical Illusion” prepublished on arXiv.org, February 20, 2019, was co-authored by Nasim Nematzadeh, David M. W. Powers and Trent Lewis.