Redes neurais profundas com redução de pesos para uso em hardwares de baixo custo aplicados à identificação de espécies invasoras em cultivares
Carregando...
Data
Autores
Orientador
Ulson, José Alfredo Covolan
Coorientador
Pós-graduação
Engenharia Elétrica - FEB 33004056087P2
Curso de graduação
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Tipo
Tese de doutorado
Direito de acesso
Acesso aberto

Resumo
Resumo (português)
O aumento na demanda por alimentos bem como a alta no preço dos
insumos para sua produção são fatores que afetam a agroindústria e a população. Neste cenário, outro desafio é o controle de pragas e doenças,
dentre essas pragas destacam-se as espécies invasoras (plantas daninhas),
nos cultivares de milho e soja, que são os grãos mais produzidos no mundo.
Atualmente, a identificação dessas plantas é um processo insalubre, laborioso e sujeito a falhas quando realizado pelo método manual. Quando o controle é feito por máquinas de maneira não seletiva, o agricultor estará beneficiando espécies resistentes tornando o manejo ineficiente, além de causar impactos ambientais pelo uso excessivo de herbicidas. Portanto, existe a necessidade de um sistema embarcado de baixo custo para a identificação de plantas daninhas que permita seu uso em tratores ou diretamente em plataformas robóticas. Com base nestes desafios, para este trabalho foi proposto e explorado o processo de identificar e localizar espécies invasoras nas culturas de milho e soja por meio de um sistema de inteligência artificial com quantização e conversão de pesos neurais em um hardware de baixo custo. Para isso fez-se uso da visão computacional associada ao aprendizado profundo (Deep Learning) no treinamento de arquiteturas de redes neurais convolucionais (Convolutional Neural Network) Yolov5 em um conjunto de imagens público da plantas daninhas prejudiciais às culturas de milho e soja. As arquiteturas foram analisadas com base nas métricas matriz de confusão, precisão, revocação, média das precisões médias (mAP) com limiar de 0,5 e média das precisões médias com limiar variando 0,5 a 0,95 (mAP 0,5:0,95). Após o treinamento fez-se uso da quantização pós treinamento para comprimir os pesos neurais para os formatos inteiro de 8 bits (int8), ponto flutuante de 16 bits (fp16), e a conversão para ONNX (Open Neural Network Exchange). Estes pesos juntamente com as imagens de validação foram transferidos para plataforma Raspberry Pi 4 model B, onde, realizou-se a avaliação dos modelos com as métricas descritas. Em seguida foi feita a inferência nas imagens do conjunto de validação. Os melhores resultados de validação foram obtidos com a rede Yolov5l com quantização em int8 com mAP 0,5 de 0,815, precisão de 0,834, revocação de 0,772 e mAP 0,5:0,95 de 0,413, na inferência a rede obteve 1 imagem inferida a cada 5,685 segundos. O modelo com menor tempo de inferência foi a Yolov5n com pesos convertidos para ONNX, onde foi obtido uma imagem inferida a cada 0,376 segundo. Neste contexto, os resultados apontam que a abordagem proposta é viável para utilização em campo em plataformas tratorizados/robotizadas. Tem-se como vantagens a possibilidade da aplicação seletiva de defensivos, poupando o meio ambiente, reduzindo custos e aumentando a eficiência na erradicação de plantas daninhas prejudiciais às culturas de milho e soja.
Resumo (inglês)
The increase in the demand for food as well as the rise in the price of
the inputs for its production are factors that affect the agro-industry and the
population. In this scenario, another challenge is the control of pests and
diseases, including invasive species (weeds) in corn and soybean cultivars,
which are the most produced grains in the world. Currently, identifying these
plants is an unhealthy, laborious and error-prone process when carried out
manually. When control is carried out by machines in a non-selective way, the farmer will be benefiting resistant species, making management inefficient and causing environmental impacts through the excessive use of herbicides. Therefore, there is a need for a low-cost embedded system for weed identification that can be used on tractors or directly on robotic platforms. Based on these challenges, this work proposed and explored the process of identifying and locating invasive plant species in corn and soybean crops using an artificial intelligence system with quantization and conversion of neural weights on low-cost hardware. For this purpose, computer vision associated with deep learning was used to train Yolov5 convolutional neural network architectures on a set of public images of weeds harmful to corn and soybean crops. The architectures were analyzed based on the following metrics: confusion matrix, precision, recall, mean average precision (mAP) with a threshold of 0.5 and mean average precision with a threshold ranging from 0.5 to 0.95 (mAP 0.5:0.95). After training, post-training quantization was used to compress the neural weights into 8-bit integer (int8) and 16-bit floating point (fp16) formats, and conversion to ONNX (Open Neural Network Exchange). These weights and the validation images were transferred to the Raspberry Pi 4 model B platform, where the models were evaluated using the metrics described. Inference was then made on the images in the validation set. The best validation results were obtained with the Yolov5l network with int8 quantization with mAP 0.5 of 0.815, precision of 0.834, recall of 0.772 and mAP 0.5:0.95 of 0.413. In inference, the network obtained 1 inferred image every 5.685 seconds. The model with the fastest inference time was Yolov5n with weights converted to ONNX, where an inferred image was obtained every 0.376 seconds. In this context, the results show that the proposed approach is viable for use in the field on tractor/robot platforms. The advantages include the possibility of selective application of defensives, which conserves the environment, reduces costs and increases efficiency in eradicating weeds that are harmful to corn and soybean crops
Descrição
Palavras-chave
Agroindústria, Plantas daninhas, Aprendizado profundo, Redes neurais convolucionais, Quantização pós treinamento
Idioma
Português
Citação
MARQUES JUNIOR, L. C. Redes neurais profundas com redução de pesos para uso em hardwares de baixo custo aplicadas à identificação de espécies invasoras em cultivares. 2023. 96 f. Tese (Doutorado em Engenharia Elétrica) - Faculdade de Engenharia, Universidade Estadual Paulista, Bauru, 2023.