Separation of plankton images from nonplankton images based on deep learning

• Ano de defesa: 2024
• Autor(a) principal: Mallqui, Diego Mauricio Mansilla
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizado por transferência; Aprendizado profundo; Classificação de imagens; Deep learning; Detritos; Detritus; Image classification; Plâncton; Plankton; Transfer learning
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/45/45134/tde-06012025-100449/

Plankton are organisms, mainly of microscopic scale, that form the base of the marine food chain. They play a crucial role in the carbon cycle. Identifying plankton species and their distribution is of great importance in research on ocean ecosystems. Modern imaging technologies are enabling the capture of large volumes of underwater images. Despite advances in image classification techniques, classifying plankton images remains challenging. Challenges include issues related to image quality, including resolution, lighting, motion blur, focal blur, low contrast, among others. Another challenge is the presence of particles, such as detritus, among the collected images. One approach to simplify the classification of plankton images would consist of first removing all non-planktonic objects from the image collection in order to facilitate the task of recognizing plankton species. The objective of this work is to develop deep learning-based methods to separate images of non-planktonic objects from those captured with the intention of recognizing plankton species. To this end, we used a set of images provided by the Oceanographic Institute of USP, containing 56,702 images from 75 classes, including the detritus, bubble and shadow classes. Five distinct convolutional neural network (CNN) architectures were trained to perform the plankton/non-plankton separation. An accuracy of approximately 95% was achieved by all models, which led us to also investigate the reasons that prevent these classifiers from surpassing this level. A careful analysis of the results provides evidence that the intrinsic visual ambiguity of the images due to their low quality, especially of blurred images, is one of the main factors that limit the separation accuracy to 95%. The analysis also indicates that a more promising way to improve plankton/non-plankton separation may be to consider a third class, i.e., plankton/non-plankton/ambiguous, in order to stimulate research aimed at better handling of these ambiguous images.