USP Develops AI Ranked Best in the World for Texture Recognition
A carpenter needs to have well-trained eyes to know by heart which wood is which just by looking at boards and scraps. This recognition can be done by looking at the colors, the shape of the darker rings and the unique grooves of each species. In a similar way, a healthcare professional studies and practices a lot until they are able to interpret an image exam accurately, identifying changes in the body that may indicate a problem.
To learn to perceive visual details like these, a modern artificial intelligence (AI) would need an image bank with millions of possibilities, a super machine with several graphics processors and many days of training. This changed when scientists from USP’s São Carlos Physics Institute (IFSC) developed Radam ( random encoding of aggregated deep activation maps ).
“The most developed AI models for processing images, with a more general view, like Google Lens , find objects, but they have difficulty when used on some type of data in which it is very difficult to see something, because they are trained to see things everyday”, says Leonardo Scabini, an IFSC researcher who led the project.
The first article about Radam was published on March 8, 2023 on the portal where it appears as the best method and, in November of the same year, in the scientific journal Pattern Recognition .
Artificial intelligence adapted by the Brazilian technique outperforms experts in some tasks. This can be used to distinguish designs without defined shapes, which to an ordinary person would be practically identical.
Just like the texture of wood, which is perceptible to the naked eye and reflects characteristics such as the roughness or smoothness of everyday surfaces, the textures visible in medical images reveal structural details that are invisible without magnification, but which reflect the properties of the cells of other microorganisms.
The photo on the left is a high-resolution microscopy of a sensor after detecting a cancerous sample and the one on the right is a healthy sample. They are images without everyday objects that, for an ordinary person, would be almost identical. A common artificial intelligence would also have difficulty noticing the differences, but Radam specializes in this type of complex pattern, depending on the purpose – Photos: Leonardo Scabini
“Sometimes, in a microscopy image of two types of cells, not even the doctor can see that one is cancerous and the other is not, because it is data that is at a very complex level or on a scale that the specialist cannot can discern nothing. But this type of artificial intelligence can do it.”
This category of algorithm can also be used in industry for inspection and quality control, in remote sensing, soil analysis, nanotechnology, environmental sciences and in biosensors for medical use.
Small-scale training
To create computer vision instruments, large companies use huge machines with multiple processors and terabytes of memory. After months, the processing of billions of information creates artificial neural networks that simulate the human brain.
Radam is unique because it synthesizes knowledge from these generic artificial intelligences and refines it for a specific problem, rather than building an intelligent architecture from scratch.
This means that it does not need thousands of data points to specialize an algorithm for each application. In addition to the scarcity of data, the cost of producing CT scans and microscopies can be quite high in many cases. Otherwise, the lack of a large volume of images available would make the resource unfeasible.
As the number of data is millions of times smaller, Radam can do in a matter of hours and on much more modest computers what conventional machine learning would take weeks.
An AI that trains other AIs
The Brazilian tool could teach the skill that some veterinarians and biologists have, such as differentiating a cat from a dog by just having an enlarged photo of a few millimeters of each one’s hair.
“It’s as if we transformed Google Lens ‘ ability to detect a leaf on a tree into the ability to detect a cell in a microscopic image with thousands of cells”, explains Leonardo.
All you need is a common laptop, without a video card and a minimum number of examples to create a program that differentiates a cancerous cell from a non-cancerous one or to identify illegal wood. “You take a microscopic image of the wood and can identify the species based on that”, explains Leonardo.
The method consists of coding what comes out of a robust and generic artificial intelligence using a small neural network. Then, an expert system is trained locally for a given challenge. This way, the machine learning process is much faster and less expensive.
Expert model
The professor emphasizes that Radam’s accuracy may be greater than that of professionals in some tested cases. “It’s a controversial subject, but there are several situations in which humans do not have the visual ability and brain to recognize the pattern in images.”
Visual texture recognition has been studied since the 1970s and in recent years has surpassed human capacity in some tasks. However, this does not mean that the technology is infallible. “She can even get a diagnosis wrong. The success rate is never 100%”, assesses professor Odemir Martinez Bruno, from IFSC, who guided the development of Radam.
According to him, a botanist could need the flower and seed to identify a plant species. The computational technique developed makes this recognition through visual patterns of a detail of the leaf or through microscopies.
“If we put two drops of blood, one from a cat and one from a dog, we cannot visually differentiate. The Radam expert system can do it”, adds the professor.
The challenge has been to interpret not only very enlarged captures, but also very reduced ones, such as space photos taken by satellites.
Constant renewal
Leonardo Scabini has explored Radam on several texture datasets, where he achieved state-of-the-art results with computers of different budgets. At the time of publication, they were the best in the world.
It turns out that, because it is an openly shared code and there are many programmers interested in improvements to this same area of computer vision, Odemir Bruno believed that another technique would surpass Radam in a few months. “The fact that the method has been ranked first in the world for a year is a very commendable thing.”
Another advantage of the system is that it can also be used in technologies that are yet to come, as Leonardo anticipates. “If someone creates a new artificial intelligence that beats our result and we apply Radam to it, it will probably improve the results. This is a great advantage, as our method is a module that can be coupled with other AIs.”
In the coming months, the researcher intends to create a second version of Radam during his stay at the Royal Institute of Technology in Stockholm, Sweden.