To enhance quality control procedures in the wood industry, the authors of the Wood Defect Detection dataset propose new automated vision-based systems. The dataset encompasses over 43,000 labeled defects found on wood surfaces, encompassing ten different types of the most frequently occurring defects. These include live_knot, dead_knot, knots_with_cracks, crack, resin, marrow, quartzite, missing_knot, blue_stain, and overgrown areas. Each image within the dataset is accompanied by a semantic map and a bounding box label, facilitating both semantic segmentation and localization tasks. Notably, all the data were directly collected from a wood production line as part of the manufacturing process. They highlight the significant variability in raw materials and the intricate manufacturing processes, leading to a diverse range of observable structural defects. These defects require assessment by trained specialists through manual processes that are laborious, subject to bias, and less efficient.

The wood industry’s manufacturing processes influence material utilization and cost efficiency. The complexity and variability of wood materials can result in various defects, affecting both mechanical properties and aesthetic value. Despite increasing automation, many companies rely on trained domain experts for quality grading, with manual inspection often limited by volume and subject to human-related factors. To address this, the authors aim to develop accurate automated systems.

The authors constructed a laboratory setup comprising a conveyor belt, light source, and line scan camera. The camera captured images at a speed of 4 m/s, synchronized with the conveyor belt’s movement. This experiment involved 250 wood veneers, each scanned at a resolution of 4,000 × 3,000 pixels. The resulting dataset contained 4,729 usable images, of which 353 had wood defects. To address the lack of extensive databases in this field, the authors performed an experiment in an industrial environment to acquire authentic data from the production line. To overcome challenges such as high-speed conveyor belts and vibrations, they developed both hardware and software solutions for high-resolution image acquisition. To ensure high-quality images at a speed of 9.6 m/s, the authors employed a trilinear line scan camera and a high-performance Camera Link frame grabber.

Annotation was performed manually, with the authors developing a customizable annotation tool to streamline the process. They created BMP files representing semantic maps of labeled defects, with each zone painted in an image corresponding to a specific defect. The annotation tool automatically generated coordinates for bounding boxes and labels. This large-scale dataset and associated labeling offer a valuable resource for advancing automated wood defect detection and quality control systems.