Automatic timber grading requires fast, accurate and consistent defect detection to support downstream solution optimization. Traditional image processing techniques based on handcrafted feature descriptors are sensitive to scanning disturbance and often require extensive domain knowledge to develop. In this master’s thesis, we develop a method to perform accurate wood defect detection in real time based on deep neural network models.
We address the problem of imbalanced defect distribution in the dataset by combining type aggregation and effective sampling, such that infrequent defects can be detected well despite inadequate annotations. On the production dataset provided by FinScan Oy, our model achieved a mean average precision of 11.63% for all 12 different wood defects and an average precision of 42.27% for dead knots, which are the most influential defects in final grading. On the RTX 2080Ti GPU, the model runs at 25 frames per second which brings a potential throughput of 250 boards per minute for the saw line.
