On October 14, NEC announced that it has developed AI technology
that captures damage appearing small in disaster images
and can also learn new types of disaster events through additional training.
The technology is the result of joint research
with the Okatani Laboratory at Tohoku University.
In the company’s verification, the number of images
requiring final human review was narrowed down to about one-ninth.
NEC aims for practical use within fiscal 2027.
What Changes: Capturing Image Details and Adding New Search Targets
The technology uses a VLM (vision-language model)
to search for on-site images that match a text description.
This development addresses two problems:
small damage being overlooked,
and earlier recognition accuracy declining through additional training.
The local region understanding technology reflects
information about the region of a target object
in the extraction of image features.
According to the company, it can grasp areas covering 10% or less of an image
with accuracy comparable to that of conventional VLMs
for objects around 30% to 40% of an image.
In the verification dataset, cases in which the affected target
covered 10% or less of the image accounted for about 60%.
The continual learning technology learns from new data
together with descriptions generated by the model before additional training,
suppressing the decline of existing recognition capabilities.
NEC says the types of events that can be searched
can be added repeatedly without preparing large amounts of past training data.
Verification Results: Narrowing 5,035 Images Down to an Average of 544
The verification used 5,035 vehicle-mounted camera images
extracted from recorded footage of the Great East Japan Earthquake
and covered 12 types of disaster events.
According to NEC, the number of images requiring final human review
averaged 544.
Assuming each image is checked visually in five seconds,
reviewing all images would take about seven hours,
compared with about 45 minutes after narrowing them down.
These times are visual review times based on the number of images
and do not represent the total time for the work, including AI processing.
Intended Use: Initial Searches and Prioritizing On-Site Checks
According to the announcement, local governments and other organizations
receive thousands to tens of thousands of on-site images
immediately after a disaster.
NEC envisions using the technology to search images with text
such as “a road blocked by landslide debris” in the initial response,
and to sort images of damage to roads and buildings
when making recovery decisions,
helping prioritize on-site checks.
LC Signal’s Perspective: Reducing What People Need to Review
The announcement shows a workflow in which AI narrows down images
and people then conduct the final review.
Support that reduces the number of images staff need to look at
and helps decide the order of on-site checks
seems likely to be a concrete starting point for adoption in disaster response.
NEC plans to combine the technology with satellite image analysis
and advance research and development
to help understand disaster situations from multiple angles.
Key Sources
-
NEC official announcement (Japanese), October 14, 2026:
NEC、現場画像に小さく写る被害を高精度に把握し、新たな被災事象にも対応できるVLM技術を開発
(NEC develops VLM technology that accurately captures damage appearing small
in on-site images and can handle new types of disaster events)
This article was prepared with AI assistance using publicly available information
as of October 15, 2026.
Announcement dates follow the dates shown on the official pages.
Verification results are based on the conditions presented by the company,
and the target for practical use is distinguished from current availability.
“LC Signal’s Perspective” reflects this publication’s interpretation
of the available information.