NEC Builds VLM Tech to Spot Small Disaster Damage and Learn New Damage Types
NEC has developed two Vision-Language Model technologies that identify small-scale damage in disaster site images and let users add new damage types without losing existing recognition ability. According to NEC, in its verification the tools cut the time to identify specific disaster conditions to less than an hour, down from about 7 hours of visual inspection.
- NEC and Tohoku University's Okatani Laboratory developed two VLM technologies for disaster image analysis.
- Small areas of 10% or less of an image are identified as accurately as conventional VLMs handle areas of about 30 to 40%, according to NEC.
- In NEC's test on 5,035 earthquake images, images needing human review fell to an average of 544 and review time to about 45 minutes versus about 7 hours.
- New damage types can be added as search criteria without large volumes of historical training data.
- NEC targets practical implementation in fiscal year 2027.
Two technologies for rescue and recovery work
NEC Corporation announced VLM-based Local Region Understanding and Continual Learning for VLMs, developed in joint research with the Okatani Laboratory at Tohoku University in Japan. NEC expects local governments and other organizations to use them in rescue operations and recovery efforts.
The first technology accurately identifies small-scale damage visible in images collected from disaster sites. The second allows new types of damage information to be added as search criteria while the VLM keeps its original recognition capabilities. Users can search large volumes of images with natural language and narrow them down to those matching a specific situation.
NEC says it targets practical implementation in fiscal year 2027.
The problem: tens of thousands of images and small details
NEC points out that natural disasters have become more severe and frequent, and that tens of thousands of on-site images often exist immediately after an event. Inspecting them by eye takes enormous effort and time for staff. VLMs that find images from text queries could let even small teams assess a situation quickly.
However, NEC identifies two weaknesses in current models. Many types of damage that matter for rescue and recovery decisions appear only as small details, occupying 10% or less of an image, and conventional VLMs sometimes overlook them. In NEC's verification dataset, the target damage occupied 10% or less of the image area in approximately 60% of cases. In addition, VLMs can struggle with previously unseen phenomena, such as region-specific events like landslides and flooding, or newly identified damage. Extra training can also reduce accuracy on events the model already recognized, a problem known as forgetting.
How the technologies work
Local Region Understanding builds region information about objects in an image into the VLM's feature extraction process. The model focuses on small areas and narrow regions, not only the whole image. NEC says it confirmed that areas accounting for 10% or less of an image can be identified with the same accuracy that conventional VLMs reach for areas of about 30 to 40%. This applies to damage such as partial structural damage to buildings or cracks in roads, and the system can generate detailed descriptions.
Continual Learning for VLMs trains on additional data and also on a diverse range of descriptions the VLM generated itself before the additional training. Learning from both sources together lets the system recognize new events while minimizing any drop in accuracy for earlier ones. Because large volumes of historical training data do not need to be prepared, search criteria can be added repeatedly later.
The research was accepted and presented at the computer vision conferences ECCV 2024 and ECCV 2026.
Verification results
NEC tested the technologies on 5,035 vehicle-mounted camera images extracted from archival footage of the Great East Japan Earthquake, covering 12 types of disaster-related events. Combining newly learned disaster events with existing features such as mountains and rivers, the number of images requiring final human verification fell from 5,035 to an average of 544, about one-ninth.
Checking every image at 5 seconds per image takes approximately 7 hours. With the technologies, NEC says review time drops to approximately 45 minutes.
What it means for disaster response
NEC describes use in two phases. In the initial response right after a disaster, text searches such as "roads blocked by landslide debris" can speed up assessment. During recovery, the tools can sort images of road and building damage to help prioritize on-site inspections. Because criteria can be added over time, users can respond to unforeseen risks such as flooding following an earthquake, and to local conditions.
Looking ahead, NEC plans to combine the technologies with analysis of wide-area observation data, including optical and synthetic aperture radar (SAR) satellite imagery, to support a multifaceted understanding of disaster conditions. It also intends to pursue research aimed at preventing disasters before they occur. The reported results come from NEC's own verification on one set of archival images.
