Contagem de laranjas utilizando rede neural convolucional
Data
Autores
Título da Revista
ISSN da Revista
Título de Volume
Resumo
Realizar a contagem de laranjas em lavouras cítricas é uma automação importante para escalabilidade e faturamento de uma lavoura. À medida que novas tecnologias são desenvolvidas, as redes neurais convolucionais tem ganhado destaque na detecção e segmentação de objetos, e por conta disso, diferentes arquiteturas são propostas em busca de proporcionar alta confiabilidade nessa tarefa. Dessa forma, propõe-se nesse trabalho a comparação de duas redes neurais convolucionais distintas, utilizando uma arquitetura de Mask R-CNN e pré-treinadas sob o dataset COCO, na tarefa específica de contagem de laranjas na laranjeira. Utilizou-se um ground truth, feito por inspeção humana, de com imagens contendo 33 resoluções distintas para o teste. Os algoritmos foram desenvolvidos no ambiente Google Colab, sendo ambos códigos disponibilizados para reprodução, estudos, melhorias e trabalhos futuros. O desempenho alcançado foi de uma acurácia média de 55,85 com Mask R-CNN e 56,04 com Detectron2.
Counting oranges in citrus crops is an important automation for scalability and billing of a tillage. As new technologies are developed, convolutional neural networks have gained prominence in object detection and segmentation, and because of this, different architectures are proposed in order to provide high reliability in this task. Thus, this work proposes the comparison of two distinct convolutional neural networks, using a Mask R-CNN architecture and pre-trained under the COCO dataset, in the specific task of counting oranges on the orange tree. A ground truth was used, made by human inspection, with images containing 33 different resolutions for the test. The algorithms were developed in the Google Colab environment, both codes being available for reproduction, studies, improvements and future works. The performance achieved was an average accuracy of 55.85% with Mask R-CNN and 56.04% with Detectron2.
Counting oranges in citrus crops is an important automation for scalability and billing of a tillage. As new technologies are developed, convolutional neural networks have gained prominence in object detection and segmentation, and because of this, different architectures are proposed in order to provide high reliability in this task. Thus, this work proposes the comparison of two distinct convolutional neural networks, using a Mask R-CNN architecture and pre-trained under the COCO dataset, in the specific task of counting oranges on the orange tree. A ground truth was used, made by human inspection, with images containing 33 different resolutions for the test. The algorithms were developed in the Google Colab environment, both codes being available for reproduction, studies, improvements and future works. The performance achieved was an average accuracy of 55.85% with Mask R-CNN and 56.04% with Detectron2.
Descrição
Palavras-chave
Citação
SILVA, Marcos Vinícius Leal da. Contagem de laranjas utilizando rede neural convolucional. 2022. Trabalho de Conclusão de Curso (Bacharelado em Engenharia Eletrônica) – Instituto Federal de Santa Catarina, Florianópolis, 2022
