Caracterización automática de especies de madera mediante técnicas de clasificación de imágenes

Resumo

Materials can be characterized determining their physical parameters. Those parameters describe the differences among them. Wood is the material studied in this thesis in its veneer form. Wood veneers are ideal for optical analysis of parameters like photometric properties, texture and fractal dimensions, which will be used to distinguish them. The main problem in wood identification is the large difference among samples from the same tree species or even from the same tree. Those differences can be greater than the differences among wood from several tree species. Pattern matching techniques based on Decision Theory will be used to solve this problem in order to identify the wood species. This work proposes firm foundations for the automatic classification of wood samples, using a low cost approach based on optical scanning with a desktop scanner combined with image processing and data mining operations, performed on a personal computer, by public domain software, Image J and Weka. The data of this thesis come from digital images of wood veneers, forming a set of properties enough for wood species identification. The influence of these parameters over the colour components of the wood images is studied. Also, the multiscaling properties of the wood images, derived from their structure, are considered. Using a novel approach inspired by Differential Contrast Interference Microscopy, pseudo-3D images and periodic patterns of the wood surface are obtained. The combination of a large number of samples and the processing of an even greater number of parameters associated with them, make necessary the reduction of the dimensions of the data matrices. Several data mining and pattern matching algorithms are used to accomplish this reduction strategy, among them are linear classifiers based on decision trees with different training methods. The results of this thesis are as good as or better than the obtained by state-ofart techniques described in the scientific literature. A high reliability has been achieved thanks to the study of a large number of species in comparison with the works of other researchers.