Image segmentation algorithm may improve efficiency 10,000 times and will be used in medical imaging

Researchers at the Massachusetts Institute of Technology recently said that they have developed a new image segmentation algorithm that can improve the efficiency of traditional segmentation algorithms by tens of thousands of times. This research will help improve the recognition accuracy of medical imaging systems and achieve continuous tracking and recognition of specific 3D objects.

When we open the window and look out, we can immediately see cars, sidewalks, pedestrians, or tall buildings in the distance. This is called visual recognition in the computer field. It is not difficult for humans, but it is a core problem that is difficult to solve for computer visual recognition technology. Because the computer does not understand the difference between two different objects, the image must be segmented to tell the computer where the boundary of each object is. The algorithm used to solve this problem is called an image segmentation algorithm.

In the image segmentation algorithm, the most primitive and traditional algorithm is to use a lot of guesswork and match and exclude through calculation. Although this algorithm can achieve the purpose, it is inefficient and takes up a lot of resources. The algorithm developed by Jason Zhang of the School of Electrical Engineering and Computer Science and John Fisher of the Computer Science and Artificial Intelligence Laboratory (CSAIL) at the Massachusetts Institute of Technology can solve this problem. They claim that the new algorithm can improve the efficiency of traditional algorithms by tens of thousands of times.

Zhang said that image segmentation is a difficult problem because there is no single correct answer. Asking 10 people is likely to yield 10 different answers. Therefore, they hope to develop an image segmentation algorithm that is similar to the way humans understand.

To achieve this goal, the algorithm of Jensen-Zhang and Fisher strikes a balance by segmenting from two aspects: first, segmenting based on color, determining the boundaries of objects according to the difference in color; second, using a fuzzy algorithm to distinguish the shapes of objects through the principle of simplification.

The experimental results show that although other researchers have adopted roughly the same approach, their original intention was to find the most suitable and unique image segmentation result, so the calculation intensity was high and the efficiency was naturally low; while the new algorithm took into account the possibility of multiple different segmentations and could perform efficient calculations with lower precision. Although there are many segmentations with low matching precision, the new algorithm can still quickly find the optimal matching result.

Anthony Izer, professor of computer engineering at Georgia Institute of Technology, said: "There are many new methods in the field of image segmentation, so it is hard to say that this segmentation method will revolutionize the entire field. But it should be affirmed that the new algorithm is very interesting, and I think it can be regarded as a milestone. This technology can be used to track objects, and it can even be used to identify tumors whose appearance changes over time. Through pattern matching, this technology can also achieve accurate recognition of objects under different angles and different lights."