A comparative study of convolutional neural networks for the problem of wood quality classification

• Grant number: 19/07248-4
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Effective date (Start): June 01, 2019
• Effective date (End): May 31, 2020
• Field of knowledge: Engineering - Production Engineering
• Principal Investigator: André Luis Debiaso Rossi
• Grantee: Gustavo das Neves Ubeda
• Host Institution: Universidade Estadual Paulista (UNESP). Campus Experimental de Itapeva. Itapeva , SP, Brazil

Abstract

Every day national companies, including the primary sector of the economy, have come up with efficient methods to qualify their production assets and thus add value to them. The classification of product quality is extremely important in order to achieve this objective, but human agents who generally perform this task are naturally subject to subjectivity and other problems arising from a repetitive process. Although in the beginning, organizations are seeking to fit the new Industry 4.0 precepts using intelligent systems, such as Machine Learning (ML) techniques. In the literature, there are several studies that investigate these techniques for the visual inspection of product quality. Recently, algorithms based on Deep Learning, such as Convolutional Neural Networks (CNNs), have achieved great success in this context for different types of materials, such as steel, textile, and wood, surpassing the predictive performance of other ML techniques. Additionally, these networks are able to perform the extraction of the most relevant characteristics of the problem automatically, that is, without a preprocessing step of the data for that purpose. However, some studies that applied the CNNs to the classification of wood quality presented inferior results in comparison to other techniques of ML. Therefore, this research project aims to compare the CNNs with traditional ML techniques for the problem of classification of wood quality from digital images. The hypothesis of this work is that by generating a greater amount of images, the CNNs will be able to surpass the predictive performance of the other algorithms investigated.