Superb AI completes first phase of LH defect classification system
The vision-language model system handles LH's roughly 4 million defect images per year and returns results around the clock.
Superb AI announced on November 2 that it has completed the first phase of building an automatic defect classification system based on a vision-language model for the Korea Land and Housing Corporation (LH), and is now running a second-phase enhancement project.
According to datanet.co.kr, the system can process LH's roughly 4 million defect images per year in real time. From a single image it classifies five items — space, part or material, defect type, work type and construction type — covering more than 18,000 classification combinations.
Results are delivered 24 hours a day in real time as natural-language sentences, together with the basis for the judgment. The system targets detection accuracy of over 90 percent and false positives and misses below 10 percent, maintained through periodic monitoring and verification.
Background
Before adopting the system, LH handled about 300,000 defect complaints per month, mainly through text and phone consultations. The existing classification scheme used the same five items and more than 18,000 combinations, with quality varying according to staff experience.
Quick answers
What does the Superb AI system do for LH?
It automatically classifies defects from a single image across five items and more than 18,000 combinations, providing results in natural-language sentences with the basis for the judgment.
How many images can the system process?
It can process LH's roughly 4 million defect images per year in real time.
What accuracy does the system target?
Detection accuracy of over 90 percent, with false positives and misses below 10 percent.