Wood texture analysis by combining the connected elements histogram and artificial neural networks

Abstract

The automatic analysis of wood texture, based on a novel concept: The Frequency Histogram of Connected Elements (FHCE) is the main contribution of this work. The FHCE represents the frequency of occurrence of a random event, which not only describes the texture's gray-level distribution, but also the existing spatial dependence within the texture. The exploitation of the FHCE's shape, alongside its wavelet transform, allows the computation of excellent features for the discrimination between sound wood and defective wood; in particular, for the really hard pattern recognition problem of detecting cracks in used wood boards. A feedforward multilayer perceptron, trained with the backpropagation algorithm, is the specific ANN classifier applied for the detection and recognition of cracks in wood boards. A large digital image database, developed after an industrial project, has been used for testing purposes, attaining a success ratio far beyond those obtained with more conventional texture analysis and segmentation techniques. © Springer-Verlag Berlin Heidelberg 2001.