Detection of defects and quality analysis of logsŠutar, Matic (Avtor) Bohak, Ciril (Mentor) Lesar, Žiga (Komentor) Wood knot segmentationRGB log imagesinstance segmentationself-supervised pretrainingmulti-view matchingKnots largely decide what a log is worth, yet they are still mostly counted by eye. Computed tomography finds them reliably but is too expensive to use where logs are harvested, while a camera is cheap but sees only the bark-covered surface, where knots are small, partly hidden and easily confused with bark. We detect knots in color images of round logs in two steps: segmenting every visible knot in each image, and matching the knots across images so that each physical knot is counted once. We release WoodKnots-600, a dataset of 600 images of 15 logs with 1171 annotated knots, including cross-view knot identities for one log. We compare nnU-Net, Ultralytics YOLO and a custom segmentation framework built on a self-supervised DINOv3 backbone under a common, log-grouped cross-validation. The custom framework performs best, with a pooled Dice score of 62.9% and an mAP50 of 38.3%, against 41.2% and 19.5% for YOLO and 30.5% and 10.2% for nnU-Net. The pretrained features drive this advantage: with supervised ImageNet weights on the same architecture, the Dice score falls to 44.8%. The model outlines the knots it finds well, but misses many small knots and knots the annotators marked as uncertain. To match knots across views, we project their masks onto a 3D reconstruction of the log and group them by spatial overlap. On annotated masks, the matching merges no distinct knots and recovers 66 of 67 knots exactly, regardless of the precise threshold.20262026-11-02 09:55:01Magistrsko delo/naloga190713sl