Medical segmentation methods in wood defect detection
Rehnbäck, Atte (2025-06-23)
Medical segmentation methods in wood defect detection
(23.06.2025)
Lataukset:
Julkaisu on tekijänoikeussäännösten alainen. Teosta voi lukea ja tulostaa henkilökohtaista käyttöä varten. Käyttö kaupallisiin tarkoituksiin on kielletty.
avoin
Julkaisun pysyvä osoite on:
https://urn.fi/URN:NBN:fi-fe2025063075893
https://urn.fi/URN:NBN:fi-fe2025063075893
Tiivistelmä
At present many quality inspection tasks in the wood processing industry are performed by production line workers if quality inspection is performed to begin with. At Maler Oy finished wooden panels are inspected and stacked for packaging by workers. This task is highly repetitive and labor intensive.
In this thesis I designed a visual wood panel inspection system based on deep learning-based methods used in medical image segmentation. Deep learning offers a versatile set of detection methods for image processing such as classification, object detection and segmentation. Segmentation has been studied extensively in the medical field and more specifically in the segmentation of tumors from magnetic resonance images. I performed a hyperparameter search in the form of an extensive literacy survey to design a robust training process for the wood feature and defect segmentation model. Tumor and wood feature segmentation are very similar problems since both use images with organic variation and imbalanced data so the results of the hyperparameter survey are transferable to the wood processing industry.
I trained and compared three networks for the wood panel feature segmentation task: U-net, DeepLabV3, and Segformer. To train the networks I collected and annotated over 6200 images of spruce interior panels with a white opaque varnish. The classes included features like cracks, edge clefts, knot holes, healthy and dry knots, pith streaks, and resin pockets. U-net performed the best of the three network architectures, so it was selected for the defect detection stage of the inspection system.
The inspection system grading component is based on contour detection and analysis. Contour analysis offers methods for extracting a feature’s dimensions and other geometric properties. Using these methods I developed a system that checks a set of predefined requirements for each wood feature class. I tested the grading system by comparing the grading results produced by the model’s test outputs with the results produced by the human made annotations. The model’s segmentations produced very similar grading results to the annotations. Therefore, it is possible to design and develop an automated solution for visual wood panel quality inspection using deep learning-based detection.
In this thesis I designed a visual wood panel inspection system based on deep learning-based methods used in medical image segmentation. Deep learning offers a versatile set of detection methods for image processing such as classification, object detection and segmentation. Segmentation has been studied extensively in the medical field and more specifically in the segmentation of tumors from magnetic resonance images. I performed a hyperparameter search in the form of an extensive literacy survey to design a robust training process for the wood feature and defect segmentation model. Tumor and wood feature segmentation are very similar problems since both use images with organic variation and imbalanced data so the results of the hyperparameter survey are transferable to the wood processing industry.
I trained and compared three networks for the wood panel feature segmentation task: U-net, DeepLabV3, and Segformer. To train the networks I collected and annotated over 6200 images of spruce interior panels with a white opaque varnish. The classes included features like cracks, edge clefts, knot holes, healthy and dry knots, pith streaks, and resin pockets. U-net performed the best of the three network architectures, so it was selected for the defect detection stage of the inspection system.
The inspection system grading component is based on contour detection and analysis. Contour analysis offers methods for extracting a feature’s dimensions and other geometric properties. Using these methods I developed a system that checks a set of predefined requirements for each wood feature class. I tested the grading system by comparing the grading results produced by the model’s test outputs with the results produced by the human made annotations. The model’s segmentations produced very similar grading results to the annotations. Therefore, it is possible to design and develop an automated solution for visual wood panel quality inspection using deep learning-based detection.