Automatic pith detection on wood boards with CNN

In the wood industry, to determine the location of the pith along the wood board is important for many reasons. For instance, based on the position of the pith, the stability and mechanical properties of the wood can be assessed thus allowing to establish the value and the possible applications of each sawn board. In particular, the pith generates the knots, imperfections that reduce the quality of the wood. However, looking only at the surfaces of the board, it is not so clear how much a knot affects the quality of the wood, but by adding the information of the position of the pith, it is possible to reconstruct the 3D pattern of the knot and actually understand how much volume it occupies within the board. The aim of this thesis is to evaluate the performance of a deep learning approach to determine the pith location. The method is based on multi-branch convolutional neural network trained by images of the board surfaces able to predict the values of x and y coordinates of the pith. The images used were generated by industrial scanners provided by Microtec Srl Gmbh and came from two different datasets. Therefore, they were initially used individually to train two separate networks and then merged into a single dataset to train a third network. The results show that this deep learning approach provides acceptable solutions.