In-domain Self-supervised Learning for Plankton Image Classification on a Budget

In the last few years, the abundance of available plank-ton images has significantly increased due to advancements in acquisition system technology. Consequently, a growing interest in automatic plankton image classification has surged. Machine learning algorithms have recently emerged to assist in the analysis of this vast quantity of data, supporting traditional manual processing. However, annotating such data is costly and demands significant time and resources, thus requiring data-efficient machine learning solutions. The typical framework for tackling this issue has been the adoption of supervised ImageNet pre-trained models, and fine-tuning them on the plankton classification downstream task. Nonetheless, self-supervised pre-training protocols may provide an effective alternative to the supervised approaches using ImageNet, while allowing the exploitation of the increasingly large amount of unanno-tated plankton data. To the best of our knowledge, no work systematically analyzes the impact of self-supervised pre-training protocols for plankton image classification. To fill this gap, in this paper, we present a thorough comparison between in-domain (plankton images) and out-of-domain (ImageNet) supervised and self-supervised pre-training, in terms of the quality of the corresponding embeddings for plankton image classification. We believe that this work may pave the way for further research in self-supervised protocols for the plankton domain, providing a valuable alternative to ImageNet, and exploiting the vast amount of unannotated available plankton images.