A CNN-based approach to measure wood quality in timber bundle images

Abstract

At present, the Smart Industry is becoming a field of great interest for many worldwide researchers since it allows to experiment and research new advanced techniques. One of the most common explored approaches in operations where image processing has already been a milestone is the use of Convolutional Neural Networks (CNN). Those networks have enhanced the current image processing algorithms, achieving an improvement in decision processes usually based on human experience, where an analytical model is not always available. This paper proposes a novel approach for measuring the number of rotted logs in timber bundles using a CNN trained on thousands of timber log images extracted from bundles. Today, the Swedish forest industry bases the selling price of timber bundles on the evaluation of a visual inspection. This operation is based on human experience to evaluate and measure timber bundles' features, which is necessary to categorize them. The proposed approach has shown promising results compared to the actual visual inspection made by operators showing an F1 score with the best CNN architecture of 0.89.