Development of a model for processing images of timber for quality assessment

I. V. Testova, M. V. Vetrova, D. S. Zaitsev

Northern (Arctic) Federal University named after M. V. Lomonosov, Arkhangelsk

Abstract: In the conditions of modern forest industry production, quality control of sawn timber is a priority task, since wood defects affect its cost, further processing and the final quality of the product. Traditional methods of visual control are often subjective, require significant time and do not always provide high accuracy. Automated systems based on computer vision and neural network algorithms can significantly increase the objectivity and accuracy of quality control of sawn timber.
Development and testing of the neural network segmentation model UNet++ are designed to automatically detect defects in images of sawn timber. The work uses a parallel algorithm for image processing, which consists of three stages:
Image preprocessing - includes noise removal, normalization, augmentation and conversion of images into tensors. Data markup was performed using the CVAT (Computer Vision Annotation Tool) platform.
Neural network processing - the UNet++ model analyzes the image and creates a mask of the detected defects. The network architecture includes a timm-mobilenetv3_large_100 encoder and a decoder with channels (256, 128, 64, 32, 16), which allows for a balance between accuracy and computational efficiency.
Post-processing of results - transfers mask contours to the original image, scales the result back and saves it for further analysis.
The developed automated segmentation method allows not only to increase the accuracy of defect detection, but also to significantly reduce the time of image processing, due to the parallel execution of all stages. This determines the feasibility of implementation at manufacturing enterprises where fast and accurate diagnostics of lumber defects is required.
The results of the study showed that the proposed approach outperforms traditional quality control methods in terms of speed and stability of operation. The developed system can be useful for woodworking plants, lumber manufacturers and furniture enterprises, where high quality of raw materials plays a key role in the production process. The research prospects lie in improving defect classification algorithms and optimizing computing resources for real-time operation.

Keywords: lumber image processing, neural networks, UNet++, image segmentation, lumber quality control, defect classification, parallel image processing.

UDC: 004.942

MSC: 91B84