Aplicação de Deep Learning na classificação de tábuas de madeira por meio de análise de imagens digitais
Carregando...
Arquivos
Data
2019-05-31
Autores
Gomes, Roger Cristhian
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
O setor madeireiro e toda sua cadeia produtiva possuem grande força e importância para a economia brasileira, representando 1,5% do produto interno bruto nacional em 2016. Toda madeira serrada deveria, idealmente, ser submetida a uma classificação para definição mais precisa do seu destino e justa de seu valor comercial. Quando essa madeira serrada é destinada ao exterior, a classificação é, na maioria das vezes, obrigatória. Nas serrarias do país que em sua maioria são pequenas e pouco automatizadas, a classificação é normalmente feita por visão humana, ou seja, um profissional faz a análise visual de cada peça e a classifica segundo algum critério. Como em todo processo que envolve capacidade humana, o erro é inerente e, nesse caso, elevado, em torno de 52%, segundo a literatura. Dada a importância do setor, a demanda de matéria prima e a necessidade crescente dessa classificação, é extremamente justificável que esse processo seja aperfeiçoado. A alternativa é a automatização, visando sobretudo o aumento no acerto dessa classificação. O objetivo deste trabalho foi desenvolver um modelo de redes neurais artificiais usando Deep Learning (DL) para a classificação automatizada de madeiras serradas de Pinus, seguindo as recomendações das normas da ABNT. O modelo aplicou Redes Neurais Convolucionais (Convolutional Neural Network - CNN), técnica muito estudada recentemente e promissora em diversas áreas, principalmente no processamento de imagens digitais e visão de máquina. Foram experimentados vários modelos, sendo o de melhor performance com acurácia de 97,50%. Comparando com outras pesquisas na mesma área concluiu-se que DL com CNN produz resultados aceitáveis na classificação de tábuas, mesmo com poucas imagens (284), diferença na variedade do Pinus (elliottii e taeda) e apresentação (madeira verde ou seca, aplainada ou não).
The timber sector and its entire production chain have great strength and importance for the Brazilian economy, representing 1.5% of the national gross domestic product in 2016. All lumber should ideally be subjected to a classification for a more precise definition of its destination and fairness of its commercial value. When this lumber is destined to the outside, classification is, in most cases, mandatory. In the country sawmills that are mostly small and little automated, the classification is usually done by human vision, that is, a professional makes the visual analysis of each piece and classifies it according to some criterion. As in any process involving human capacity, the error is inherent and, in this case, high, around 52%, according to the literature. Given the importance of the industry, the demand for raw materials and the growing need for such classification, it is extremely justifiable that this process is improved. The alternative is automation, aiming in particular to increase the accuracy of this classification. The objective of this work was to develop a model of artificial neural networks using Deep Learning (DL) for the automated classification of Pinus sawn timber, following the recommendations of ABNT standards. The model applied Convolutional Neural Network (CNN), a very recently studied and promising technique in several areas, mainly in digital image processing and machine vision. Several models were tried, being the one of better performance with accuracy of 97.50%. It was concluded that DL with CNN produces acceptable results in the classification of boards, even with few images (284), difference in the Pinus variety (elliottii and taeda) and presentation (green or dry wood, planed or not).
The timber sector and its entire production chain have great strength and importance for the Brazilian economy, representing 1.5% of the national gross domestic product in 2016. All lumber should ideally be subjected to a classification for a more precise definition of its destination and fairness of its commercial value. When this lumber is destined to the outside, classification is, in most cases, mandatory. In the country sawmills that are mostly small and little automated, the classification is usually done by human vision, that is, a professional makes the visual analysis of each piece and classifies it according to some criterion. As in any process involving human capacity, the error is inherent and, in this case, high, around 52%, according to the literature. Given the importance of the industry, the demand for raw materials and the growing need for such classification, it is extremely justifiable that this process is improved. The alternative is automation, aiming in particular to increase the accuracy of this classification. The objective of this work was to develop a model of artificial neural networks using Deep Learning (DL) for the automated classification of Pinus sawn timber, following the recommendations of ABNT standards. The model applied Convolutional Neural Network (CNN), a very recently studied and promising technique in several areas, mainly in digital image processing and machine vision. Several models were tried, being the one of better performance with accuracy of 97.50%. It was concluded that DL with CNN produces acceptable results in the classification of boards, even with few images (284), difference in the Pinus variety (elliottii and taeda) and presentation (green or dry wood, planed or not).
Descrição
Palavras-chave
Aprendizado profundo, Automatização, Pinus, Redes Neurais Convolucionais, Automation, Convolutional Neural Networks, Deep learning, Pinus