Improving Knot Segmentation Using Deep Learning Techniques

In the context of Computed Tomography scanning of logs, accurate detection of knots is key for delivering a successful product. Reliable detection of knots in the sapwood is hard with traditional computer vision techniques, because of the different density conditions between sapwood and heartwood. The advancements provided by deep learning in the field of semantic image segmentation kick-started a new way of approaching such problems: deep neural networks can be trained on large amounts of labelled data and successfully employed in production environments to improve the performances on knot detection. Adapting state-of-the-art network architectures and using more than 10.000 labelled knots from pine and spruce logs, we were able to develop a new two-step approach for identifying knots in CT scans of logs with unprecedented accuracy while at the same time satisfying the time constraints that a real-time industrial application needs. The first step runs on the log’s axis, while the second runs on each candidate knot’s axis. False positives from the first step are very rare (even with dry/dried logs), so no computational power is wasted for non-existing knots. Using this approach, we are able to see the internal defects of a log in real time in the production chain without having to cut it first, therefore being able to optimize even more the output of the chain on each client’s requirements.