An investigation on deep learning approaches for diatoms classification

Diatoms are one of the largest groups of microalgae present in marine, freshwater and transitional environments and their reactivity to environmental changes makes them suitable to be employed as biomarkers for monitoring tasks. Anyway, their presence in a large number of species makes it arduous to perform diatoms taxonomy during monitoring tasks considering that, to date, analysis is conducted by marine biologists on the basis of their own experience and, hence, in a subjective way. Hence, the need for automatic and objective methodologies for the identification and classification of diatoms samples rises. Research efforts in the field of Computer Vision led to a plethora of highly effective deep learning strategies surpassing human capabilities for image classification, as showed in the recent Imagenet challenge editions where they were initially introduced. Despite the very promising results of the proposed solutions, the difficulty arises to determine which technique is most suitable among them for real tasks and in particular for diatoms classification. This work proposes an end-to-end pipeline for automatic recognition of diatoms, acquired by means of holographic microscopy in water samples, employing deep learning techniques. In particular the most recently introduced Convolution Neural Networks (CNNs) architectures have been deeply investigated and compared in order to highlight the pros and cons of each of them. Moreover, in order to feed the CNNs training stages with a suitable amount of labeled data, a strategy to build a synthetic dataset, starting from a single image per class available from commercial glass slides specifically prepared for taxonomy purposes, is introduced. Besides, models ensembling strategies, in order to improve the single model scores, have been exploited. Finally, the proposed approach has been validated employing a dataset built up of holographic images of diatoms sampled in natural water bodies.