Computer Vision for Virtual Sawing and Timber Tracing

Abstract

Process optimisation in the sawmill industry is, as in other major industries, an ongoing endeavour that aims to derive maximum benefit from advancements in measurement technology and process control. Raw material undergoes a significant transformation during the sawing process. Therefore, different imaging and processing techniques are required at different stages of production. The challenging tasks of virtual sawing and timber tracing can be solved by bridging this gap between the stages. Virtual sawing is used to predict various properties of the sawing output before sawing the log, potentially allowing for the optimisation of sawing parameters. Timber tracing refers to the problem of tracing the origin of a board to the log it was sawn from. This information can be used to avoid faulty products at an earlier production step or to ensure the legality and sustainability of the product origin.

This study focuses on the development of a novel, unified framework for the solution of both the virtual sawing and timber tracing tasks using point clouds of logs and images of boards. A method for the conversion of raw log point clouds into a concise and informative 2D heightmap of the log’s surface is proposed. The virtual sawing task is resolved by using a novel fuzzy volumetric model for the internal knot distribution created based on log heightmaps. The results of the virtual sawing are used to approximate knot locations on the resulting boards. Generative adversarial models are used to create photorealistic prediction of the visual appearance of boards. An extension of the volumetric model with annual rings is proposed in order to simulate the fibre patterns on the virtual boards. A solution for the timber tracing problem based on the longitudinal distribution of knot clusters is proposed and evaluated. The method uses encoder-decoder convolutional networks to transform log heightmaps and board images into the modality representing knot cluster locations.

This study uses data collected from a real sawmill environment to develop and evaluate novel methods for virtual sawing and timber tracing. The data is collected using laser range scanners for logs and digital cameras for boards. The experimental results are provided and discussed. The results indicate that even though only the surface information is used, enough information can be extracted to make an estimate of the log’s internal structure. A comparison of the predicted virtual boards with the real ones is presented. The evaluation results of the proposed timber tracing method are discussed. The results suggest that the method is more suitable for low-quality logs. However, accuracy can be improved by using additional information, such as log and board sizes. Overall, the accuracy of the knot location prediction is limited, since only surface information is used. In the future, additional data can be used to improve the methods based on deep learning. The algorithm can also be potentially augmented with information from other sensors and extended to solve more tasks, such as sawing angle optimisation.