Virtual sawing using generative adversarial networks

Daniel Batrakhanov, Fedor Zolotarev, Tuomas Eerola, Lasse Lensu, Heikki Kälviäinen

Abstract: The global trend towards digitalization allows building new innovative solutions to optimize manufacturing. In particular, the highly competitive sawmill industry is not an exception. The industry always depends on efficient raw material utilization, and thus, exploration of the internal structure of wooden logs is an important feature in the timber conversion process. One common approach to a comprehensive internal wood structure examination is virtual sawing, that is predicting the outcome of sawing process based on log measurements. In this work, the suitability of generative adversarial network (GAN) based image-to-image translation methods for solving the virtual sawing problem is studied. We propose an extended virtual sawing framework where a GAN based Pix2Pix model is utilized to create photorealistic images of boards. Furthermore, we propose a novel evaluation protocol for assessing the GAN generated virtual boards by utilizing timber defect segmentation. In the experimental part of the work, we show that the defects (knots) on virtual boards are detectable and their locations correspond to those in real boards.