Math-Based AI Surpasses Super-Resolution Limitations
Expected Applications in Criminal Investigations and Smart Cities

A new artificial intelligence (AI) technology developed by a Korean research team has raised the global standard for accurately reading blurred car license plates captured by CCTV at night or in adverse weather conditions. This achievement is drawing attention for its potential real-world applications, as it goes beyond simply enhancing image quality to significantly improving actual recognition performance.


On March 18, Korea University announced that a research team led by Professor Oh Seungsang of the Department of Mathematics had won first place in the "Low-Resolution License Plate Recognition (LRLPR) Competition," hosted by the International Conference on Pattern Recognition (ICPR 2026).

Low-resolution License Plate Restoration Comparison Table. Provided by Korea University

Low-resolution License Plate Restoration Comparison Table. Provided by Korea University

View original image

A total of 537 researchers from around the world participated in the competition, putting their technological skills to the test. The Korea University team, comprised of graduate students Jung Hyuk Jeong, Bae Subin, and Yoo Jeongbin, along with undergraduate students Kim Donggeon and Seo Euikwan, achieved an accuracy rate of 82%—correctly identifying all seven characters on license plate images that were difficult to discern. Only four teams among all the participants surpassed the 80% accuracy mark.


Not Just "Visible" but "Readable"...AI Recognition Performance Elevated


Existing optical character recognition (OCR) technologies have long faced limitations in low-resolution environments. When there is severe compression damage or light smearing, the strokes and boundaries of characters are often distorted, making them difficult to read.


To address this, super-resolution technologies have been utilized. However, while these may make images appear clearer to the human eye, they have had limitations in restoring the information necessary for AI to recognize characters.


The Korea University research team addressed this issue by applying a "Teacher-Student Knowledge Distillation Learning Framework," which combines super-resolution and character recognition technologies. This method transfers the recognition capabilities of a "teacher model" trained on high-resolution images to a "student model" operating in low-resolution environments. Additionally, a mathematical optimization technique that integrates three different AI models was used to further enhance recognition rates.


This technology is particularly significant for its high potential for real-world application. License plates can now be recognized even under challenging conditions such as nighttime, rain, or fog, allowing not only for hit-and-run investigations but also for deployment across public safety networks, including intelligent transportation systems (ITS) and automated enforcement systems in smart cities.


Professor Oh Seungsang said, "Our laboratory's strength lies in approaching AI based on mathematical theory," and added, "We will continue to produce research results that combine mathematics and AI, focusing on both practicality and competitiveness."



The award ceremony for this competition is scheduled to be held at the main ICPR 2026 conference in Lyon, France, this August.


This content was produced with the assistance of AI translation services.

© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.

Today’s Briefing