Banca de DEFESA: JOÃO CÉLIO LUNA DE CARVALHO

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : JOÃO CÉLIO LUNA DE CARVALHO
DATE: 29/07/2025
TIME: 09:00
LOCAL: Google meet
TITLE:

Development of a Computer Vision System for the Selection of Agricultural Products Using Convolutional Neural Networks.

KEY WORDS:

Processamento de imagens; aprendizado de máquinas; visão computacional, maturação do tomate; injúrias em laranja.

PAGES: 95
BIG AREA: Ciências Agrárias
AREA: Engenharia Agrícola
SUBÁREA: Máquinas e Implementos Agrícolas
SUMMARY:

The demand for high-quality horticultural products, properly classified and selected, is increasing in line with global food production and population growth. To support this activity, the application of artificial intelligence—particularly through neural networks and image processing—has become increasingly common. In general, the allocation of agricultural products by the industry depends on quality standards, yet equipment focused on classification and selection remains scarce in Brazil. Convolutional Neural Networks (CNNs), a subset of deep neural networks, are specifically designed for image processing in a faster and more efficient manner. Therefore, the objective of this project was to develop a device capable of using a computer vision system associated with CNNs to classify agricultural products based on quality attributes such as the presence of epidermal injuries, ripeness level, and variety—whether under controlled lighting conditions or in the field. The image selection chamber consisted of a Raspberry Pi 5 with 8 GB of RAM connected to a dedicated camera module, an LCD monitor, and LED lamps for lighting. The software for CNN training and data classification was stored on the internal memory, provided by a 64 GB Micro SD card. Six different datasets were built under varying environmental and lighting conditions to test the robustness of the algorithm. Two CNNs with established architectures—AlexNet and MobileNet—were trained from scratch (without pre-trained weights), and their performances were compared with a custom-developed CNN, created during the project, named OakMoonNet. Performance comparison was conducted using Accuracy, Precision, Recall (Sensitivity), and F1-Score metrics for each neural network. OakMoonNet outperformed the other models in all tests, both in terms of execution time and evaluation metrics. The developed device, along with its software and user interface, also proved to be user-friendly and intuitive, enabling the broader use of CNNs even by individuals with limited technical knowledge.

COMMITTEE MEMBERS:
Presidente - 2161955 - ANDERSON GOMIDE COSTA
Interna - 2942088 - JULIANA LOBO PAES
Externo à Instituição - EMANOEL DI TARSO DOS SANTOS SOUSA - UFRPE