Scale Sensitivity of Textural Features

Prevailing surface material recognition methods are based on textural features but most of these features are very sensitive to scale variations and the recognition accuracy significantly declines with scale incompatibility between visual material measurements used for learning and unknown materials to be recognized. This effect of mutual incompatibility between training and testing visual material measurements scale on the recognition accuracy is investigated for leading textural features and verified on a wood database, which contains veneers from sixty-six varied European and exotic wood species. The results show that the presented textural features, which are illumination invariants extracted from a generative multispectral Markovian texture representation, outperform the most common alternatives, such as Local Binary Patterns, Gabor features, or histogram-based approaches.