Name: JAIRO LUCAS DE MORAES
Publication date: 14/07/2023
Examining board:
Name | Role |
---|---|
ALBERTO FERREIRA DE SOUZA | Presidente |
CLAUDINE SANTOS BADUE | Coorientador |
ELIAS SILVA DE OLIVEIRA | Examinador Interno |
FABIO LUIZ PARTELLI | Examinador Interno |
FELIPE MAIA GALVÃO FRANÇA | Examinador Externo |
Pages
Summary: The cultivation of papaya plays an essential role in the economy of several countries, serving as a significant
source of income and job creation, particularly in developing nations. Within the fruit industry, papaya holds significant
relevance, being grown in over 60 countries, including Brazil, which stands out as the second-largest producer of this
fruit. However, papaya is a delicate and climacteric fruit, leading to substantial post-harvest losses ranging from 30%
to 40% of the total production. In this regard, early detection and accurate classification of diseases and fruit damage
play a crucial role in quality control and mitigating losses throughout the production chain. Currently, papaya quality
control is conducted manually, requiring exhaustive and repetitive efforts, along with specialized knowledge that may
not always be available to farmers in remote areas or small fruit processing facilities. Moreover, manual assessment
is subject to evaluator subjectivity, including varying levels of expertise and psychological state, resulting in imprecise
interpretations of the diseases present in the fruits. Given this scenario, there is an urgent need to implement
autonomous or semi-autonomous systems that aid in papaya quality control, encompassing disease detection,
ripeness assessment, and identification of physical damage. These technological solutions are highly desirable for
mitigating losses in the industry, offering a more efficient, precise, and reliable approach to ensuring fruit quality and
maximizing papaya fruit production. In this thesis, we propose a comprehensive solution that encompasses the
creation of a unique dataset in the literature, the development of a mobile application, and the implementation of
novel Convolutional Neural Network (CNN) structures utilizing the Convolutional Block Attention Module (CBAM) and
frameworks for evaluating new detector models. Our dataset comprises over 23,000 examples of eight types of injuries
that affect papaya fruits, along with healthy fruit samples. The developed detector leverages cutting-edge resources
and achieves a new state-of-the-art in disease detection in papaya fruits, with an average precision of 86%. This
performance significantly surpasses that of human experts, who achieved an average precision of 67%. Finally, we
optimized the structure and weights of our detector to ensure high-performance inference on mobile devices, creating
a robust mobile application capable of running on standard smartphones. This application can detect diseases in
papaya fruits at a rate of up to 8 frames per second without requiring any additional resources.