Deep learning algorithm for the automatic detection of the pith position on wood boards

reservedThis thesis introduces a deep learning model designed for the grading of wood boards. In par- ticular, the focus is on the prediction on wood boards of the pith of the log from which they originate, which, as sustained in the literature, represents a critical factor in assessing the board quality. The data consists of images of the surfaces of wood boards of the Douglas fir species, obtained from industrial optical scanners. For this task, a deep multi-branch 2D Convolu- tional Neural Network (CNN) architecture proved to perform effectively. Model evaluation shows considerable results, achieving an R2 score of 0.94 and 0.97 respectively for the x- and y- coordinates of the position of the pith. The model’s predictive accuracy is further highlighted by the visual evaluation of the predictions on boards of different species. The fact that the av- erage inference time for each board is around 200 ms and the resources needed are limited to a NVIDIA GeForce GTX 980 GPU with just above 700 MBytes to deploy, makes the model represent a solution aligned with the industry standards. This work contributes successfully to advance in the assessment of one of the main aspects involved in the grading of wood boards in particular underlining the potential of deep learning techniques.This thesis introduces a deep learning model designed for the grading of wood boards. In par- ticular, the focus is on the prediction on wood boards of the pith of the log from which they originate, which, as sustained in the literature, represents a critical factor in assessing the board quality. The data consists of images of the surfaces of wood boards of the Douglas fir species, obtained from industrial optical scanners. For this task, a deep multi-branch 2D Convolu- tional Neural Network (CNN) architecture proved to perform effectively. Model evaluation shows considerable results, achieving an R2 score of 0.94 and 0.97 respectively for the x- and y- coordinates of the position of the pith. The model’s predictive accuracy is further highlighted by the visual evaluation of the predictions on boards of different species. The fact that the av- erage inference time for each board is around 200 ms and the resources needed are limited to a NVIDIA GeForce GTX 980 GPU with just above 700 MBytes to deploy, makes the model represent a solution aligned with the industry standards. This work contributes successfully to advance in the assessment of one of the main aspects involved in the grading of wood boards in particular underlining the potential of deep learning techniques