Localization and Classification of Defects on Natural Wood Surfaces

Automating the industrial quality control of natural products such as natural wooden surfaces is challenging. The core problem in this domain of anomaly detection is that each surface is composed of statistical features and, therefore, unique. Hence, locating defects and defining the normality on such surfaces becomes a difficult task. To address this problem, we first briefly present existing industrial anomaly detection methods, from which we select two state-of-the-art methods and evaluate them with respect to defect localization on natural wooden surfaces. In contrast, we also evaluate saliency maps and compare them to the selected state-of-the-art methods. Furthermore, we present a new dataset that addresses the challenge of anomaly detection in real wood products. As a result of the investigation, it is found that a successful inspection of the images is possible with state-of-the-art methods. However, localization works much better with salience maps. In this way, product quality can be improved through more precise localization of defect locations, and statistics can be obtained to eliminate possible systematic defects caused by the production process. Additionally, this paper provides valuable insights into the effectiveness of state-of-the-art methodologies.