Developed VLM technology that accurately identifies small-scale damage in on-site images and can respond to new types of disaster damage

NEC developed VLM localized-region understanding technology that accurately identifies, from on-site images collected by municipalities during disasters, even damage-related events that appear small in the images. In parallel, it developed continual-learning technology that can add new types of disaster damage while maintaining the VLM's recognition capabilities. The results were announced as the outcome of joint research with the Okatani Laboratory at Tohoku University.

Conventional VLMs had difficulty identifying cases in which damage appeared small in on-site images. The localized-region understanding technology incorporates regional information and is said to enable highly accurate identification even of small-scale damage covering 10% or less of the area. The continual-learning technology incorporates existing explanatory text as well as additional training data, enabling recognition to be expanded while limiting forgetting when new events are added.

Tests showed that the technology can reduce missed instances of small-scale damage in on-site images and shorten the time required to identify specific disaster conditions from approximately 7 hours to just under approximately 1 hour. It is expected to support searches for blocked roads and prioritization during recovery decisions, as well as responses to changes in disaster risk and local conditions.

NEC plans to use the technology to accelerate municipalities' disaster response. Going forward, it also plans to pursue research and development aimed at multifaceted assessment and prevention by combining ground images with wide-area observation data such as satellite imagery.