日本語

Inspection in the dark, with fully offline AI.

An edge AI system that runs on a single existing PC in the field and does not send images outside.

In development (a PoC at a manufacturing site is planned)

PoC and deployment inquiries

An original image of the inside of a dark factory, and an image in which it has been brightened and the equipment is outlined with boxes (both are concept images made with an image-generation AI). “Status: Local / Offline” and “AI Model: Ready” are displayed at the top.
A prototype screen in development. It shows the low-light original image on the left, and the results of image enhancement and object detection on the right. The two images on the screen are concept images made with an image-generation AI.
A concept image of the inside of a factory with the lights down. The outlines of the machine tools can barely be made out.The low-light original image (concept image)A concept image of the same factory, brightened. Machine tools and tables are outlined with boxes and labeled “Anomaly Detected (98%)” and “Component A (95%)”.The result of image enhancement and object detection (concept image)
The two images shown on the prototype screen in development. Both are concept images made with an image-generation AI at the planning stage, not actual results of processing.

In the field, inspection in the dark still relies on the human eye

In visual inspections in dark places, anomalies and deterioration are more easily overlooked. Those oversights can lead to serious accidents and occupational injuries.

Image analysis by AI helps with this. Even so, one of the reasons companies hesitate to adopt it is the risk of information leaks that comes with sending images to the cloud.

From input to result

One existing PC in the field
Image and instructionImage enhancementObject detectionLocal LLMDisplay of results
Images do not leave this outline.

Beacon is an edge AI system that corrects images taken in dark places in the field so that they can be seen, and catches anomalies automatically. It combines low-light image enhancement and object detection with a local LLM. It runs on a single existing PC in the field and analyzes images without sending them outside.

  1. Image and instruction inputA natural-language prompt is added to a low-light image.
  2. Autonomous parallel processingThe system runs image enhancement and object detection at the same time.
  3. Immediate display of resultsThe results are shown on the screen in near real time.

The image enhancement method is published in papers and code

Beacon’s image enhancement uses a Retinex-based method by its developer, Toma Okugawa. XCR, which is fast and preserves structure, was published as a preprint in 2025, and its successor, SD-Retinex, in 2026.

SD-Retinex is a method designed around the derivative of the sigmoid function. It requires no training and runs on a CPU alone. Among the training-free CPU methods compared, it showed the best results.

From the SD-Retinex paper. Evaluated on the LOL dataset (eval15), on a CPU (AMD Ryzen 7 8840U).

PSNR (higher is better)
SD-Retinex16.47 dB
XCR9.83 dB
SSR8.99 dB
Processing time per image (shorter is faster)
SD-Retinex18.4 ms (54 FPS)
XCR24.4 ms
LIME31.8 ms
SSR151 ms
MSR604 ms
  • The color difference ΔE (lower is better) was 15.9; all the other methods compared were 30.9 or higher.
  • Without training, it outperformed the zero-reference deep learning methods Zero-DCE and SCI in both PSNR and ΔE.

Awards

Tongali Business Plan Contest 2026 (June 2026)

  • NICT Award
  • Supporter Award: JR Central Award
  • Supporter Award: Beyond Next Ventures Award
  • Tongali Award, 5th place

Entry title: “Beacon: A Proposal for a Fully Closed-Network Edge AI System Using Low-Light Image Enhancement and a Local LLM”

Team: Toma Okugawa, Uryu Den, Keito Inoshita

Starting at manufacturing sites, then on to disaster sites and medical care

  • Short term (up to 1 year)PoC at manufacturing sites
  • Medium term (2–3 years)Expansion to disaster and accident sites
  • Long term (3+ years)On to medical care, inspection and night security

PoC and deployment inquiriesTo the Fieldlight contact form