Research Highlights

**This post does not wish to be quoted, redistributed, or reported by external media or third-party press.

Professor Kangseok Joo's research lab (Taewoo Kim, integrated M.S.–Ph.D. program), wins Gold Prize and Special Prize at Samsung Display Industry-Academia Collaboration Paper Competition

Professor Kangseok-Ju's research team in the Department of Electronic Engineering (Taewoo Kim, an integrated M.S.–Ph.D. program student)published a paper that 2026 won the gold prize at the Samsung Display Industry-Academia Collaboration Paper Awards. Additionally, Professor Kangseok-Ju's research lab also received a special award rank given to the lab with the most outstanding papers 3for the year, achieving a feat of winning consecutively. To publicize the research achievements of industry-academia collaborations 2018starting from the first year, the Samsung Display Industry-Academia Collaboration Paper Awards are held annually 10in November 19at the Samsung Display Giheung Campus Samsung Display Researchin room R514.

The title of the paper selected for the gold award is 'Anyshot: A Unified Framework for Full and Few-Shot Anomaly Detection' and , it proposes an industrial anomaly detection (anomaly detection) framework that operates robustly even in a data imbalance environment in actual manufacturing sites. The research team pointed out that existing anomaly detection methods assume that normal learning data exists equally for all product groups, but in actual production lines, 'inter-category shot imbalance (inter-category shot imbalance)' occurs, where product groups with abundant data and those with insufficient data coexist due to reasons such as production volume and collection costs. Additionally, existing rebalancing methods, such as downsampling or synthetic data augmentation, have limitations in that they discard actual data or distort the normal distribution.

This paper proposes a method to overcome these limitations by training a single model directly on the imbalanced original data distribution without artificially manipulating the data. 'Anyshot'and proposes, It consists of three key components.. First,, 'the class-balanced weighted sampler equalizes the expected frequency of each product group within a mini-batch without modifying the stored dataset. (class-balanced weighted sampler)'Second,. the dual-branch, 'architecture applies a transformer-based reconstruction decoder to product groups with abundant data and a normalized memory bank to those with scarce data, (dual-branch)' providing an explicit normality prior even with minimal data., to product groups with scarce data, a normalized memory bank (memory bank)is applied to 1-shot provide an explicit normality prior even with a minimal amount of data.. Third,, 'the hybrid scoring (hybrid scoring)' module fuses the anomaly maps from the two branches using element-wise minimum, (element-wise minimum)enabling stable detection without separate hyperparameter tuning or labeled validation data., enabling stable detection without separate hyperparameter tuning or labeled validation data..

The research team evaluated MVTec AD, VisA, Real-IAD datasets with five random mixed-shot (mixed-shot) settings using Anyshot. As a result, Anyshot achieved the highest and most stable image-level anomaly detection performance compared to 10 state-of-the-art models, with the greatest performance advantage in the few-shot product group where the impact of imbalance is the greatest, especially in the few-shot product group. The proposed Anyshot technology can utilize imbalanced data from actual manufacturing sites without a complex data rebalancing process, so it is expected to greatly contribute to the automation of quality inspection in various industries, such as displays, and other fields.

3 years in a row, winning a special award, which once again proved the outstanding research achievements of Professor Seokju Kang, who "Anyshot redefined data imbalance as a learning problem that the model should absorb on its own, rather than a preprocessing problem that needs to be removed before learning, which is significant" said, "it will play an important role in actual industrial inspection environments where it is difficult to secure normal data, such as new products or small-batch production products", adding that. Additionally, , the first author of this study, Taewoo Kim, a researcher, 1st emphasized that "this study will be the foundation for anomaly detection technology that can be directly applied to the data environment of actual manufacturing lines", stressing that.

‣proposed Anyshot framework

‣Qualitative comparison results of existing and proposed methodologies

‣Paper Title: Anyshot: A Unified Framework for Full and Few-Shot Anomaly Detection

‣Author Information: Taewoo Kim, integrated M.S.–Ph.D. program (the 1st author), Seokju Kang, Professor (corresponding author)

‣Other Information: None