Recognition of wood surface defects with near infrared spectroscopy and machine vision

Abstract

To improve the accuracy in recognizing defects on wood surfaces, a method fusing near infrared spectroscopy (NIR) and machine vision was examined. Larix gmelinii was selected as the raw material, and the experiments focused on the ability of the model to sort defects into four types: live knots, dead knots, pinholes, and cracks. Sample images were taken using an industrial camera, and a morphological algorithm was applied to locate the position of the defects. A portable near infrared spectrometer (900–1800 nm) collected the spectra of these positions. In addition, principal component analysis was utilized on these variables from spectral information and principal component vectors were extracted as the inputs of the model. The results show that a back propagation neural network model exhibited better discrimination accuracy of 92.7% for the training set and 92.0% for the test set. The research reveals that the NIR fusing machine vision is a feasible tool for detecting defects on board surfaces.