Defesa de Qualificação de Doutorado do discente Pedro Henrique Nascimento Castro

Defesa de Qualificação de Doutorado do discente Pedro Henrique Nascimento Castro, dia 12/04/2024 às 08:30.

Título: EXPLORING MACHINE LEARNING TECHNIQUES FOR HOP VARIETY CLASSIFICATION AND OUT-OF-DISTRIBUTION DETECTION

Abstract: Humulus lupulus L., commonly known as hops, is a vine plant integral to brewing, imparting distinct flavors, bitterness, and aromas to beer, as well as contributing to foam stabilization. Its antimicrobial and antioxidant properties are not only pivotal in beer preservation but have also been explored for food preservation. Additionally, hops find applications in the cosmetics industry and are recognized for their health benefits. In Brazil, hop production is in its early stages. Despite growth in the sector, a considerable part of the national crop is still in an incipient phase. As a result, few nurseries have a certificate of hop origin, leading to products that may have different characteristics than expected. Thus, accurately distinguishing the variety to which the plant belongs is essential for producers. With over 250 cataloged hop varieties, each possessing unique characteristics, the identification of hop variety based on its acids and essential oils is typically achieved through techniques like chromatography, mass spectrometry, capillary electrophoresis, and nuclear magnetic resonance. However, these methods involve costly and complex equipment, often inaccessible to many hop producers. In this work, we pursue the classification of hop varieties through leaf imagery, a more cost-effective and accessible method. We developed a database comprising 1,592 images of 12 popular hop varieties in Southeast Brazil. Initially, traditional machine learning techniques were employed to establish a baseline for classification. Subsequently, using deep learning, we achieved over 95% accuracy in classification through a multi-view method that combines leaf detection in the image with the full image analysis. Additionally, we contribute to the literature by presenting a novel regularization approach for deep learning models called Simultaneous Learning and a technique for model evaluation named Layer Correlation.

Data: 12 de abril de 2024 Horário: 8:30

Local: Sala de Seminários DECOM - ICEB III

Banca: Prof. Dr. Eduardo Luz, Prof. Dr. Gladston Juliano Prates Moreira (UFOP); Prof. Dr. Pedro Henrique Lopes (UFOP), Prof. Dr. David Menotti (UFPR); Profa. Dra. Sandra Ávila (UNICAMP)