%0 Conference Proceedings
%4 sid.inpe.br/sibgrapi/2021/09.15.01.13
%2 sid.inpe.br/sibgrapi/2021/09.15.01.13.52
%T Um Estudo Comparativo de Redes Convolucionais Profundas para Detecção de Insetos em Imagens
%D 2021
%A Domênico, Jéssica Regina Di,
%A Lau, Douglas,
%A Ribeiro, Daniel Delfini,
%A Rieder, Rafael,
%A Júnior, Telmo De Cesaro,
%@affiliation Instituto Federal de Educação Sul-rio-grandense (IFSul)
%@affiliation Embrapa Trigo
%@affiliation Instituto Federal de Educação Sul-rio-grandense (IFSul)
%@affiliation Universidade de Passo Fundo (UPF)
%@affiliation Instituto Federal de Educação Sul-rio-grandense (IFSul)
%E Paiva, Afonso,
%E Menotti, David,
%E Baranoski, Gladimir V. G.,
%E Proença, Hugo Pedro,
%E Junior, Antonio Lopes Apolinario,
%E Papa, João Paulo,
%E Pagliosa, Paulo,
%E dos Santos, Thiago Oliveira,
%E e Sá, Asla Medeiros,
%E da Silveira, Thiago Lopes Trugillo,
%E Brazil, Emilio Vital,
%E Ponti, Moacir A.,
%E Fernandes, Leandro A. F.,
%E Avila, Sandra,
%B Conference on Graphics, Patterns and Images, 34 (SIBGRAPI)
%C Gramado, RS, Brazil (virtual)
%8 18-22 Oct. 2021
%I Sociedade Brasileira de Computação
%J Porto Alegre
%S Proceedings
%K yolo mask rcnn cnn cnn aphids.
%X This work presents a comparative study between two deep convolutional network models in tasks of identification and counting of insects in digital images, considering aphids (Hemiptera: Aphididae) and parasitoids (Hymenoptera: Aphelinidae and Braconidae, Aphidiinae). In this case study, each image can contain hundreds of specimens, debris, overlaps, and other insects with similar morphology, making the detection process difficult. In this sense, we compared the results obtained by the InsectCV system, which was based on Mask R-CNN, in terms of training time, inference, and precision, with a new model, trained with the DarkNet network. Using grayscale images with smaller dimensions, processing via GPU, and a one-stage convolutional network, it is possible to reduce the computational cost and increase the precision in the object detection task. Based on the 580 images used to validate the proposed model, it was possible to obtain a mean Average Precision of 79.9\%.
%@language pt
%3 paper.pdf