• Back
• Next

日本語

Humans have the ability to sense texture in order to estimate the conditions of a surface or the quality of a material. This enables us to sense rich texture from, for example, the beautiful glitter of a precious metal or the colored, translucent line of light from a jewel. Throughout history, humans have always prized high quality textures and the intricate reflections and transmissions of light on the surfaces of objects. Against this backdrop, the effort to understand the human brain's processing of texture sensing has actively been conducted across various academic fields such as engineering, psychology and neuroscience.

A reflective material like a mirror or polished metal has specular reflections on its surface. Transparent materials like glass or ice allow light to pass through and light refracts inside them. For these two types of material, the image visible on the surface of the material may change in significant and complicated ways depending on the material's surroundings. Therefore, there are countless possible permutations, and it has been largely unclear how humans distinguish between them.

This led the research team to conduct psychophysical tests to discover how accurately humans distinguish between reflective and transparent materials. The team also verified how accurate convolutional neural networks (CNNs) could be when making the same distinction. The test showed that a human being could distinguish between reflective and transparent materials at 78% accuracy while the CNN could make the same distinction at 94% accuracy, which is considerably more accurate than humans.

From these results, we can see that the CNN’s accuracy is great enough to create potential for industrial applications where it may replace human observers. However, the question we really want to solve is, "How do humans distinguish between the two different types of material?" The research team concluded that it would be hard to outperform humans in identifying the image clues that humans perceive from the structure or behavior of a model.

Therefore, we tuned the CNN not only to answer correctly as humans do, but also to dare to "make mistakes as humans do," and verified what is used as a cue based on the structure of the model and its similarity to humans. The results showed that a relatively shallow model with three layers of CNN convolution structure was the most similar to that of humans, suggesting that the model may use image changes that appear on the top of the object as a cue. These findings support the insights into human texture sensing reported in previous studies.

This research became the first to successfully structure a model that enables image computing and distinguishes between reflective and transparent materials while imitating a human being’s correct and incorrect answers. Applying this model may make it possible to distinguish between materials and reproduce textures based on summarized data without having to use all of the data in an image. In other words, we can expect that there will be applications for this model in technologies achieving the highly accurate reproduction of textures at low cost.