Hanwha Vision has claimed first place at one of the world's most prestigious international AI competitions after developing a breakthrough continual learning technology that overcomes a longstanding limitation in artificial intelligence. The achievement is expected to accelerate the company's development of next-generation security systems that combine video and audio intelligence.
Hanwha Vision announced on Monday that its joint research team with Professor Hong-Gook Kim's laboratory at the Gwangju Institute of Science and Technology (GIST) won first place in the DCASE 2026 Challenge, an internationally recognized competition organized by the Institute of Electrical and Electronics Engineers (IEEE) that evaluates state-of-the-art acoustic AI technologies.
This year's competition attracted 135 teams from universities and research organizations worldwide.
The Hanwha Vision-GIST team secured the top position in the Domain-Agnostic Incremental Learning for Audio Classification track, outperforming 20 competing teams by demonstrating superior performance in continual learning for environmental sound recognition.
The challenge required AI systems to accurately classify ten categories of everyday and safety-related sounds—including baby cries, barking dogs and fire alarms—even when recordings were captured using different devices and in varying acoustic environments.
The winning technology addresses one of artificial intelligence's most difficult challenges: catastrophic forgetting, a phenomenon in which AI models lose previously learned knowledge when trained on new data.
Traditional audio AI systems often experience significant declines in recognition accuracy after learning new environments. For example, an AI model trained to recognize urban sounds may struggle to retain that knowledge after being updated with audio collected in airports or other unfamiliar settings.
Such limitations have hindered wider adoption of AI in applications where operating environments frequently change, including smart surveillance, industrial safety and smart city infrastructure.
To overcome this problem, the research team developed a DeepInversion-based Generative Replay technique that enables AI models to reconstruct characteristics of previously learned audio without requiring the original training data. This approach allows the system to preserve earlier knowledge while continuously learning new information.
The researchers also incorporated an ensemble methodology that combines predictions from multiple AI models, further improving recognition accuracy and system stability.
The final ensemble system achieved an average classification accuracy of 79.62%, significantly outperforming the competition average of approximately 68%. Notably, the team's ensemble model and three additional standalone systems occupied the top four positions in the official rankings, underscoring the robustness of the technology.
Hanwha Vision plans to integrate the continual learning technology into its AI-powered video surveillance portfolio to create multimodal security solutions capable of analyzing both visual and acoustic information simultaneously.
By combining audio recognition with intelligent video analytics, future security platforms will be able to detect abnormal situations more quickly and accurately across a broader range of environments, the company said.
"This achievement demonstrates that Hanwha Vision's AI research capabilities are among the best in the world," said Lim Jung-eun, Head of the AI Research Center at Hanwha Vision. "We will continue advancing continual learning technologies to develop next-generation AI security solutions capable of adapting autonomously to changing environments."
The joint research team will formally present its findings and receive the championship award at the DCASE 2026 Workshop, scheduled to be held in Boston, Massachusetts, on Oct. 28–29.
The research was conducted through a joint industry-academia collaboration between Hanwha Vision and GIST, with support from Hanwha Vision's internal R&D program and a project funded by South Korea's Ministry of Trade, Industry and Energy.
