Professor Oh Ki-yong from Hanyang University School of Mechanical Engineering and his research team developed the “Rotational Multipyramid network, RoMP Net,” which is an image recognition technology with highly enhanced recognition accuracy and versatility, according to Hanyang University on September 10. The RoMP Net is a “One-stage detector” that can provide real-time image recognition thus is expected to be widely applied to our everyday lives, such as autonomous mobility and precision diagnosis.

Artificial intelligence-based image recognition is one of the core technologies of the 4th Industrial Revolution, which is used mainly for △autonomous driving △content meta information provision △medical image analysis. So far, deep neural networks for image recognition such as SSD (Single-shot multibox detector) and YOLO (You Only Look Once) had poor prediction accuracy as the background environment information was also included during the learning when the background environment was included inside the bounding box. In particular, when recognizing images in an environment that was not learned, the recognition accuracy largely decreases. Furthermore, because shallow, deep neural network structures were used, complicated or small objects were sometimes not recognized.

Professor Oh’s team enhanced the prediction accuracy by minimizing the background environment included in the learning by installing a rotational bounding box to the RoMP Net. Also, they stacked the pyramid structured deep neural network in a multi-level to create deep neural networks that can recognize small or complicated objects.

The result of testing the RoMP Net on an actual inspection system for autonomous drones used when testing transmission lines showed that the recognition accuracy increased more than 10% and versatility more than 200% compared to the previous deep neural network used for image recognition.

Professor Oh said, “The key to securing accuracy and versatility was optimizing the complex system of the deep neural network, which can extract only the necessary elements in an object. This research took the artificial intelligence-based image recognition technology one step ahead and secured a source technology for next-generation autonomous flight and driving.”

This research, recognized for its innovativeness, was selected as the cover paper for the international academic journal in computer science and artificial intelligence International Journal of Intelligent System(IF=10.312) on September 1.

Meanwhile, this research was supported by the Electrical Engineering & Science Research Center of Korean Electric Power Corporation.