U-NET-based deep learning for automated detection of lathe checks in homogeneous wood veneers

Abstract

Automated detection of lathe checks in wood veneers presents significant challenges due to their variability and the natural properties of wood. This study explores the use of two convolutional neural networks (U-Net architecture) to enhance the precision and efficiency of lathe checks detection in poplar veneers. The approach involves sequential application of two U-Nets: the first for detecting lathe checks through semantic segmentation, and the second for refining these predictions by connecting fragmented lathe checks. Post-processing techniques are applied to denoise the mappings and extract precise lathe check characteristics. The first U-Net demonstrated strong performance in predicting lathe check presence, with precision and recall scores of 0.822 and 0.835, respectively. The second U-Net refined predictions by linking disjointed segments, improving the overall lathe checks mapping process. Comparative analysis with manual methods revealed comparable or superior performance of the automated approach, especially for shallow lathe checks. The results highlight the potential of the proposed method for efficient and reliable lathe check detection in wood veneers.