Wood species recognition in open set framework using fuzzy classifier and generalised basic probability assignment

It seems impossible and unnecessary to recognise all tree species within a wood classification system. In practice, the special tree species in special areas need to be recognised and simultaneously other species rejected. These special species are of interest and are included in our training dataset, whereas the other species rejected are not interesting to us and are not in the training dataset. To correctly and simultaneously classify known wood species and reject unknown species, a novel wood species recognition system in an open set framework is proposed. A Flame-NIR spectrometer was used to obtain the nearinfrared (NIR) spectral curve of wood cross sections. The spectral feature dimension reduction was performed using a principal component analysis (PCA) processing. The fuzzy if-then classifier was revised to extend its use from a closed set to an open set. A novel generalised basic probability assignment (GBPA) is proposed based on the product of membership degree values and the grade of certainty of a fuzzy rule. The comparison experimental results on both balanced and imbalanced datasets indicated that our proposed scheme outperforms the classification schemes using one class classifier in open set and the state-of-the-art wood classification schemes in closed set.