Comparison of feature extraction methods applied to wood quality classification

• Grant number: 18/11852-1
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: December 01, 2018
• End date: November 30, 2019
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: André Luis Debiaso Rossi
• Grantee: Natália Fernandes Soares
• Host Institution: Universidade Estadual Paulista (UNESP). Campus Experimental de Itapeva. Itapeva , SP, Brazil

Abstract

The success and survival of industries are increasingly dependent on producing products and services of high quality. Recently, this is occurring even for small and medium-sized suppliers of raw material, such as timber industries. In this area, the classification of wood quality is a decisive factor to add value to the product and satisfy the requirements of the customers. This task is usually performed by human specialists, who are exposed to different types of subjectivity, such as their experience and knowledge, and the variations that may occur due to fatigue or other difficulties. Therefore, the automation of this process is of great interest to these industries. A current project at UNESP, Itapeva, investigates the use of Machine Learning (ML) techniques for automatic visual inspection of wood boards. The ML area studies the development of systems capable of improving their predictive performance from previous experiences. For the visual inspection problem, this experience is obtained from images of wood boards captured by cameras and classified by a human expert. A prototype that integrates these techniques into low-cost cameras has already been tested successfully in a timber industry in the region. Despite the improvement achieved in comparison with human operators, the performance of the ML techniques employed for this task depends on the characteristics that will be used to describe the images. The process of obtaining relevant image information is called feature extraction. Therefore, the present research project aims to investigate image feature extraction methods for the problem of wood quality classification in order to improve the predictive performance of the models generated by the ML techniques. This research will be performed using 374 images of wood boards collected of an industry of the region of the campus. These images were previously classified by an expert considering three levels of quality. Each feature extraction method will produce a data set where each predictive attribute corresponds to an image characteristic and the target attribute corresponds to the quality level. The models generated by the different ML algorithms will be compared in relation to the predictive performance of different baselines. In addition, these models can provide some evidence of which characteristics were important to the task under analysis.