A paper developed within the ENCIRCLE project has been accepted for presentation at the 34th European Signal Processing Conference (EUSIPCO 2026), to be held in Bruges, Belgium, from 31 August to 4 September 2026. EUSIPCO is one of Europe’s premier conferences on signal processing, bringing together leading researchers and practitioners for a high-quality programme of tutorials, keynotes, poster and lecture sessions.
The accepted paper, titled LMM-assisted Subclass-Aware Zero-Shot Industrial Defect Detection, is authored by Maria Tzelepi, Nikolaos Dimitriou, Christos Gkogkos and Dimitrios Tzovaras from CERTH, the ENCIRCLE project coordinator.
The work addresses industrial defect detection in manufacturing environments where labeled defective data is scarce. The authors propose a novel LMM-assisted Subclass-Aware method that improves the performance of CLIP-based Zero-Shot Anomaly Detection (ZSAD) models. An LMM specialised for anomaly detection is used to extract defect-type descriptions from an auxiliary dataset, which are then used to compute defect-type center embeddings. A subclass objective is introduced during training to maximise similarity between defective image embeddings and their closest defect-type center, promoting cross-category defect-type discrimination and delivering significant performance improvements over the state of the art.
This second paper acceptance at a major international conference within the same period reflects the growing scientific output of the ENCIRCLE project and the quality of research being produced by the consortium.
More about EUSIPCO 2026 at https://eusipco2026.org/
ENCIRCLE Publications https://encircle-project.eu/scientific_publications/