Automatic insect count in trap images using deep learning
Data
2022-07-29
Autores
Título da revista
ISSN da revista
Título do Volume
Editora
Resumo
As pragas de insetos são a principal causa de perda de produtividade
e qualidade nas culturas em todo o mundo. O percevejo e a traça da uva são
exemplos de duas das pragas mais significativas que afetam o arroz e as vinhas,
respetivamente. As armadilhas para insetos estão entre as soluções mais adequadas
para monitoramento e contagem influenciando a seleção e dosagem do pesticida
a ser aplicado no controlo de pragas. No entanto, a monitorização e contagem
baseiam-se na visita frequente de técnicos ao local e são suportadas por métodos de
contagem ineficientes, sendo uma tarefa exigente e demorada. A gestão integrada de
pragas foi desenvolvida para melhorar a gestão de pragas de insetos, reduzir o uso
excessivo de pesticidas e aumentar a qualidade e o rendimento das culturas. Com o
aprimoramento das tecnologias de inteligência artificial, diversas aplicações surgiram
no contexto agrícola, incluindo detecção automática, monitorização e contagem de
insetos.
Este estudo propõe a contagem automática de insetos em armadilhas usando algoritmos de deep learning. Foram utilizados três bases de dados diferentes, Pest24, Bedbug e Grape moth. Pest24 é um conjunto de dados público com uma grande diversidade de insetos. Os conjuntos de dados Bedbug e Grape moth são conjuntos de dados privados fornecidos pelo mySense, uma plataforma de agricultura de precisão desenvolvida e gerida por investigadores da UTAD.
Nossa metodologia foi dividida em duas partes, o uso de detetores de um estágio
(YOLOv5) e de dois estágios (Faster R-CNN). Para a YOLOv5, realizamos um
total de 8 experimentos, nos quais foram analisados o impacto do transfer learning
e o hyperparameter tuning. Para a Pest24, obtivemos um desempenho superior ao
relatado no estado da arte (72,1% de mAP), com o método YOLOv5. Em seguida,
usando os hiperparâmetros e weights obtidos do conjunto de dados Pest24, treinamos
os conjuntos de dados Bedbug e Grape moth. Os melhores resultados para o dataset
do Bedbug foram obtidos com o YOLOv5 com tranfer learning com um AP de 95,9%.
O melhor resultado foi obtido com YOLOv5 usando o ajuste de hiperparâmetros da
Pest24 com AP de 91,5% para o Grape moth. Para o detector de dois estágios,
realizamos 8 experimentos, nos quais foram combinados o uso de CNN (VGG ou
Inception ResNet V2) e a aplicação de otimização de âncora, ajustando a âncora
ao tamanho dos insetos. A otimização da âncora mostrou-se eficaz no conjunto de
dados Grape moth, atingindo um AP de 80%, mas para o conjunto de dados da
Grape moth, o método que teve o melhor desempenho foi sem a otimização âncora
com um AP de 78.3%.
Como resultado, YOLOv5 foi a melhor arquitetura, apresentando apenas dificuldades
de detecção em imagens com elevado número de insetos. Além disso, foi possível
verificar que a escala relativa dos insetos é o fator que mais afeta a tarefa de detecção e que a otimização da âncora pode melhorar os resultados de detecção.
Insect pests are the leading cause of yield loss and quality in crops worldwide. The bedbug and the grape moth are examples of two the most significant pests a_ecting rice and vineyards. Insect traps are among the most appropriate solution for monitoring and counting influencing the selection and dosage of the pesticide to be applied for pest control. However, the counting and monitoring operations are based on the frequent visit of technicians to the site and are supported by inefficient counting methods, which is a challenging and time-consuming task. Integrated pest management was developed to improve insect pest management, reduce the overuse of pesticides, and increase the quality and yield of crops. With the improvements of artificial intelligence technologies, several applications have emerged in the agricultural context, including automatic detection, monitoring and counting of insects. This study proposes the automatic counting of insects in traps using deep learning algorithms. Three different databases, Pest24, Bedbug and Grape moth were used. Pest24 is a public dataset with a great diversity of insects. The Bedbug and the Grape moth datasets are private datasets provided by mySense, a precision agriculture platform developed and managed by researchers from the UTAD. Our methodology was divided into two parts, the use of one-stage (YOLOv5) and two-stage (Faster R-CNN) detectors. For the one-stage detector, we carried out a total of 8 experiments, in which the impact of transfer learning and hyperparameter tuning were analyzed. First, for Pest24 dataset we obtained a performance superior to that reported in state of the art, with the YOLOv5 method with standard hyperparameters, with an mAP of 72.1%. Then, using the hyperparameters and weights obtained from the Pest24 dataset, we trained the Bedbug and Grape moth datasets. The best results for the Bedbug dataset were obtained with the YOLOv5 with transfer learning with an AP of 95.9%. The best result was obtained with YOLOv5 using the hyperparameters tuning of Pest24 with an AP of 91.5% for the Grape moth. For the two-stage detector, we performed 8 experiments, in which the use of CNN (VGG and Inception ResNet V2) and the application of anchor optimization or not (adjusting to the size of the insects) were combined. The anchor optimization proved to be effective in the Bedbug dataset, reaching an AP of 80%, but for the Grape moth dataset, the method that performed the best was without the anchor optimization with an AP of 78.3%. As a result, YOLOv5 was the best architecture, presenting only detection difficulties in images with a high number of insects. Furthermore, it was possible to verify that the relative scale of the insects is the factor that most affects the detection task and that the anchor optimization can improve the detection results.
Insect pests are the leading cause of yield loss and quality in crops worldwide. The bedbug and the grape moth are examples of two the most significant pests a_ecting rice and vineyards. Insect traps are among the most appropriate solution for monitoring and counting influencing the selection and dosage of the pesticide to be applied for pest control. However, the counting and monitoring operations are based on the frequent visit of technicians to the site and are supported by inefficient counting methods, which is a challenging and time-consuming task. Integrated pest management was developed to improve insect pest management, reduce the overuse of pesticides, and increase the quality and yield of crops. With the improvements of artificial intelligence technologies, several applications have emerged in the agricultural context, including automatic detection, monitoring and counting of insects. This study proposes the automatic counting of insects in traps using deep learning algorithms. Three different databases, Pest24, Bedbug and Grape moth were used. Pest24 is a public dataset with a great diversity of insects. The Bedbug and the Grape moth datasets are private datasets provided by mySense, a precision agriculture platform developed and managed by researchers from the UTAD. Our methodology was divided into two parts, the use of one-stage (YOLOv5) and two-stage (Faster R-CNN) detectors. For the one-stage detector, we carried out a total of 8 experiments, in which the impact of transfer learning and hyperparameter tuning were analyzed. First, for Pest24 dataset we obtained a performance superior to that reported in state of the art, with the YOLOv5 method with standard hyperparameters, with an mAP of 72.1%. Then, using the hyperparameters and weights obtained from the Pest24 dataset, we trained the Bedbug and Grape moth datasets. The best results for the Bedbug dataset were obtained with the YOLOv5 with transfer learning with an AP of 95.9%. The best result was obtained with YOLOv5 using the hyperparameters tuning of Pest24 with an AP of 91.5% for the Grape moth. For the two-stage detector, we performed 8 experiments, in which the use of CNN (VGG and Inception ResNet V2) and the application of anchor optimization or not (adjusting to the size of the insects) were combined. The anchor optimization proved to be effective in the Bedbug dataset, reaching an AP of 80%, but for the Grape moth dataset, the method that performed the best was without the anchor optimization with an AP of 78.3%. As a result, YOLOv5 was the best architecture, presenting only detection difficulties in images with a high number of insects. Furthermore, it was possible to verify that the relative scale of the insects is the factor that most affects the detection task and that the anchor optimization can improve the detection results.
Descrição
Mestrado Integrado em Engenharia Eletrotécnica e de Computadores
Palavras-chave
Deep learning , Contagem de insetos