Visão computacional para identificação de espécies lenhosas em campo

Abstract

Anatomical identification of forestry species is an important tool for control and supervision of timber trade, mainly because it enables the protection of vulnerable species. The recent increase in the demands of the international timber market and the full evolution of technological areas such as machine learning and machine vision have encouraged the development of intelligent and automatic species identification systems based on wood images. In this work, we sought to develop and evaluate a computer vision system capable of identifying species with a smartphone to acquire images of manually polished samples with knives. Three image banks were built, the first containing 2,000 images of 21 species and the second containing 30,200 images of 40 species. The third, also with 40 species, was formed with 32,271 images obtained with superficial moistened wood samples. Three image descriptors were evaluated: gray level coocurrence matrix, local binary patterns and gray level run length matrix. Different resolution settings and gray levels of the images were also evaluated. A total of 49 statistical classifiers were developed using support vector machines and evaluated in nested cross validations. The great majority of the classifiers tested presented accuracies higher than 90%, local binary patterns performed better than the other image descriptors and the sample wetting did not improve the classification performance. The proposed system was able to reach 99.36% accuracy, surpassing the results obtained in all works of the consulted literature. The simple methodology used in this work, associated with this high accuracy, makes evident the potential for the automated identification of wood with machine vision system under field conditions