Adapting object detection models to underwater environments presents significant challenges due to variable light conditions and limited data availability. This work introduces a novel approach to enhance underwater object detection by synthetically generating realistic underwater images using diffusion models. Initially, the model is trained on a dataset of real underwater images, after which it automatically generates new images that reflect the underwater domain. Automated inference is achieved using the proposed verification block, which checks the suitability of the produced image based on cosine similarity. This approach significantly expands the size of underwater datasets, which may enhance the performance of object detection models in challenging underwater environments.

Алшавареб Б. (науч. рук. Колюбин С.А.) Synthetic dataset generation for underwater object detection using diffusion models // Сборник тезисов докладов конгресса молодых ученых. Электронное издание. – СПб: Университет ИТМО, [2025]. URL: https://kmu.itmo.ru/digests/article/14794