An open-source web-based recognition system: automated classification, detection, and explainability with a case study on pollen grains

• Ano de defesa: 2025
• Autor(a) principal: Furtado, Emanuel Ferreira
• 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: Aplicação web; Aprendizado profundo; Artificial intelligence; Deep learning; Inteligência artificial; Pollen recognition; Reconhecimento de pólen; Web application
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/76/76135/tde-18092025-083228/

Pollen grains play a fundamental role in plant reproduction and in maintaining biodiversity. Understanding pollen biology provides valuable insights into pollination mechanisms, plant evolution, and ecological interactions. Currently, pollen detection systems rely on manual collection and analysis processes, which are time-consuming, labor-intensive, and prone to errors. The identification and classification of pollen grains remain a major challenge, especially considering the high diversity of plant species and the morphological variability of pollen grains. Furthermore, manual counting and classification are tasks that demand considerable time and may lead to inconsistencies due to human error. In this dissertation, we present an open-source web application developed using deep learning techniques for the automated classification and detection of pollen grains in microscopic images. The main goal of the application is to simplify the process for users without technical expertise, reducing the effort required compared to traditional manual methods. To achieve this, both traditional machine learning approaches and deep learning techniques were systematically explored to improve the accuracy of the results. Given the need and importance of transparency and reliability in artificial intelligence (AI) models, explainable AI (xAI) techniques — such as gradient-based methods, perturbation techniques, and feature visualization — were applied to various models, including DenseNet, EfficientNet, ResNet, and hybrid CNN-Transformer architectures. The results demonstrated that deep learning models significantly outperform traditional methods, offering greater accuracy, speed, and reliability. Additionally, the developed web platform proved to be intuitive, facilitating real-time pollen recognition and enhancing the interpretability of the models used. Beyond representing an advancement in automated pollen recognition, this approach also establishes a solid foundation for future research in interdisciplinary AI applications.