Learning to tell the rough from the smooth
Distinguishing between textured surfaces is surprisingly tricky for machines, but a computer model that runs on a PC has taught itself how.
Detecting damage on the surface of a historical artifact. Quality control of fabrics on production lines. Decoding satellite images with greater speed and precision. Each of these tasks hinge on being able to quickly and accurately identify surface texture.
Humans excel at this intuitively. We can glance at images or 3D objects and instantly tell smooth from rough. Machines, however, typically struggle. Most existing systems require powerful computers and large labelled datasets tailored to a specific texture-identification task.
Now, a team of researchers from the Computer Science Department at Khalifa University has developed software that can train itself to identify and differentiate surface textures without human annotation. Running on a standard PC equipped with a commercially available graphics processing unit (GPU), the system can distinguish between smooth regions and those with hair-like patterns seen on 3D scans of statues and pottery.
“Segmenting textures helps in understanding surface properties, relief patterns, restoration of cultural artifacts and robotic vision.”
Iyyakutti Iyappan Ganapathi
“The goal was to segment texture from non-texture regions in 3D artifacts without relying on manual annotations,” explained KU’s Iyyakutti Iyappan Ganapathi. “Labelling each facet is laborious, so we designed an unsupervised method that learns from the surface geometry directly.”
The model uses neural networks to assess the 3D surfaces, which are divided into a mesh of hundreds, or thousands, of individual facets. Rather than relying on human-labelled data, the model initially assigns each facet a label of smooth or textured, based on patterns in the data. Then it refines these labels by finding and reinforcing patterns in the data, effectively learning where it has made mistakes and correcting them.
Importantly, the software model is fast. It can process up to half a million facets in just a few minutes, depending on the spatial resolution.
The approach could be used for many different applications, Ganapathi explains. “Segmenting textures helps in understanding surface properties, relief patterns, restoration of cultural artifacts and robotic vision. Being unsupervised, it scales easily to large datasets and unseen surfaces.”
As a next step, the team plans to extend the framework to more complex texture classification tasks including finer grained textures.
Reference
Ganapathi, I.I., Dharejo, F.A., Javed, J., Ali, S.S. & Werghi, N. Unsupervised dual transformer learning for 3-D textured surface segmentation IEEE Trans. Neural Netw. Learn. Syst, 36 (3), 5020-5031, 2025. | Article
