A Few-Shot Learning Based Environmental Microorganism Image Classification Method

Because of the difficulty of collecting Environmental Microorganisms (EMs) and the limited availability of publicly accessible environmental microbial image datasets, the progress of related scientific research has been hindered. The research of EMs is also in trouble because of the lack of samples. To this end, we establish the EMDS-6 dataset that can be used for deep learning. This dataset includes 21 types of EMs, with each type represented by 40 original images and 40 corresponding Ground Truth images, for a total of 1680 EMs images. Nevertheless, image classification of small EMs datasets on convolutional neural networks continues to yield suboptimal outcomes. In order to further solve the small data problem, we did a study on Few-shot Learning (FSL) on this basis. Furthermore, leveraging self-supervision, our approach enhanced the accuracy to 57.40<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> across 21 EMs classes, demonstrating the potential of FSL and Self-Supervision in the image classification of compact EMs datasets.