Lomin Participates as Platinum Sponsor of KCCV 2026, Unveils Strategy Built on Grounding and Confidence Estimation

The performance of document AI should be measured by the level of real business automation it delivers, rather than by accuracy alone. The argument rests on reducing the human review burden through technology that isolates only the results most likely to contain errors.

Lomin announced on the 7th that it participated as a Platinum Sponsor of KCCV 2026, the annual conference of the Korean Computer Vision Society, held at BEXCO in Busan from the 3rd to the 5th, where the company delivered an industry session presentation and operated an exhibition booth.

In the industry session, held under the title "Document Understanding: Vision-Language Challenges in Real-World Environments," Lomin explained enterprise document AI outcomes through the lens of the business automation rate — the share of documents that can be processed without a person having to review or correct them.

Im Bi, Chief Technology Officer of Lomin, stressed the point: "Even if document recognition accuracy reaches 99%, if you cannot identify which documents contain the errors, every document has to be checked again by a person. In that case, the model's average accuracy is high, but the actual business automation rate can remain at 0%."

The CTO identified grounding and confidence estimation as the core technologies for raising that rate. Grounding shows where in the source document a piece of AI-extracted information came from. Confidence estimation selects only the results with a high probability of error and routes them to a human for review.

"In confidence estimation, what matters is not designing the model to assign high scores to correct results, but designing it to assign low scores to results that are actually wrong," he explained. "That allows enterprises to progressively expand the set of documents that require no human verification, within an acceptable risk threshold."

During the event, Lomin also unveiled its Document AI Agent platform at its exhibition booth. The platform supports the entire document processing cycle, including document parsing and classification, information extraction and comparison, de-identification of sensitive information, and AI training.

The company additionally hosted an on-site meetup to recruit AI researchers at the master's and doctoral level as well as specialized research personnel. This year's KCCV featured roughly 170 papers and drew some 1,300 participants, the largest turnout in the conference's history.

CTO Im, who delivered the presentation, is the first author of "EDSR," a paper on image super-resolution. The paper has been cited more than 15,000 times according to Google Scholar, and the work earned him the Sang-Uk Lee Best Paper Award at KCCV in 2022.

"What enterprises expect from document AI is not simply high accuracy, but how reliably it can automate actual business work," Im said. "Building on the experience we have accumulated in the field, we will continue to advance document AI technology optimized for enterprise environments."