Understanding multimodal timber matching networks via activation maps

Abstract

The use of digital technologies in timber tracing from the raw material to the end product, i.e., from logs to boards, provides benefits for the sawmilling industry such as better product quality prediction, efficient process control, and optimization. However, identification of boards from logs instantly becomes challenging after actual sawing, and that is why an automatic method for board identification was developed in previous studies. This thesis focuses on analyzing the multimodal encoder-decoder networks developed for timber matching to better understand what the network actually learns during the training process. The proposed solution is to compute activation maps and visualize the features learned by the neural network. Activations are further analyzed to estimate their ability to provide useful information based on the data collected from a sawmill. Finally, the use of activation maps for defect detection is studied. The results demonstrate that activation maps can be used to detect knots on both RGB images of boards and laser surface scans of logs in a weakly supervised manner.