Unsupervised anomaly detection from wooden boards using autoencoders

Abstract

For wood processing in the sawmill industry, quality of the raw material in every step affects the production efficiency. Defects in the sawn timber, such as wane, knots, cracks, watermarks, fungal damage, insect defects, can decrease its strength, durability, and usefulness, reducing the economic value. Image-based methods for anomaly detection can be used in quality-controlled manufacturing in sawmills in order to reduce the time for inspecting sawn timber. In this study, an autoencoder neural network is studied for implementing anomaly detection for wooden boards. Based on the experiments, detection performance of the selected architecture for the autoencoder was not high enough for practical application of the method.