A deep clustering framework for underwater image recognition

Underwater image recognition plays a crucial role in assessing the health status of marine ecosystems. By utilizing underwater cameras and image recognition technology, researchers can monitor the biodiversity, population numbers, growth status, and overall structure and functionality of ecosystems in the ocean. However, the problem of marine ecology assessment always occurs in dynamic and open environments, and discoveries of unknown new species are often made. Existing works which applied classification methods directly may not address this situation well. Therefore, unsupervised learning is needed to cluster these newly emerged species. However, due to strong noise interference in underwater images, clustering the unlabeled samples directly is difficult. To address this issue, we propose a two-stage training framework that can learn discriminative knowledge from labeled data for clustering new classes. Its core idea is to utilize pseudo-labeling to train the model, and then strengthens the capability of clustering by leveraging the consistency between the labeled and unlabeled data. Furthermore, contrastive learning is also used to optimize the model's representation in the embedding space. Experimental results on the WildFish dataset of over 5000 species verified the effectiveness of the proposed method in open-set underwater image recognition.