목재 단면 중심부 자동 검출 딥러닝 모델 비교 연구

Automatic Pith Detection in Tree Cross-Section Images Using Deep Learning Tzu-I Liao Oregon State University Corvallis, Oregon zack.liao@oregonstate.edu Mahmoud Fakhry Oregon State University Corvallis, Oregon fakhryk@oregonstate.edu Jibin Yesudas Varghese Oregon State University Corvallis, Oregon jibinye@oregonstate.edu Abstract Pith detection in tree cross-sections is essential for forestry and wood quality analysis but remains a man- ual, error-prone task. This study evaluates deep learning models—YOLOv9, U-Net, Swin Transformer, DeepLabV3, and Mask R-CNN—to automate the process efficiently. A dataset of 582 labeled images was dynamically augmented to improve generalization. Swin Transformer achieved the highest accuracy (0.94), excelling in fine segmentation. YOLOv9 performed well for bounding box detection but struggled with boundary precision. U-Net was effective for structured patterns, while DeepLabV3 captured multi-scale features with slight boundary imprecision. Mask R-CNN initially underper- formed due to overlapping detections, but applying Non- Maximum Suppression (NMS) improved its IoU from 0.45 to 0.80. Generalizability was next tested using an oak dataset of 11 images from Oregon State University’s Tree Ring Lab. Additionally, for exploratory analysis purposes, an addi- tional dataset of 64 labeled tree cross-sections was used to train the worst-performing model to see if this would improve its performance generalizing to the unseen oak dataset. Key challenges included tensor mismatches and bound- ary inconsistencies, addressed through hyperparameter tuning and augmentation. Our results highlight deep learn- ing’s potential for tree cross-section pith detection, with model choice depending on dataset characteristics and ap- plication needs. 1. Introduction 1.1. Background Pith detection in tree cross-sections is essential for forestry, dendrochronology, and wood manufacturing, pro- viding valuable insights into tree growth patterns, wood quality, structural integrity, and historical climate condi- tions. The pith, located at the geometric center of the tree, acts as a crucial reference point for analyzing annual growth rings, detecting anomalies such as decay or reaction wood, and estimating mechanical properties that are vital for struc- tural lumber grading. Accurate pith localization significantly impacts indus- trial wood processing, particularly in sawmills, where pre- cise pith detection facilitates optimal log cutting strategies, reduces wood waste, and improves the structural integrity and economic value of wood products. Furthermore, in den- drochronology, precise pith localization is pivotal for recon- structing past environmental and climatic conditions, as tree ring analyses depend heavily on accurately identifying the initial growth rings starting from the pith. Historically, pith detection has been performed manually by skilled analysts using calipers, magnification tools, and image-processing techniques such as thresholding and edge detection. However, manual methods suffer from subjec- tivity, inconsistencies between analysts, and inefficiencies when processing large datasets. With the increasing de- mand for precision in forestry and wood engineering, there is a need for automated solutions that can deliver consis- tent, scalable, and accurate pith detection across diverse tree species and conditions. Recent advancements in deep learning have demon- strated remarkable success in image-based segmenta- tion and object detection tasks. Deep learning models, particularly convolutional neural networks (CNNs) and transformer-based architectures, have the potential to sur- pass traditional image processing methods in accuracy and efficiency. By leveraging these advancements, automated deep learning models can detect pith locations with higher consistency, reduced manual effort, and improved scalabil- ity across large datasets. 1.2. Motivation The complexity of tree cross-sections poses a signifi- cant challenge for pith detection. Variations in wood grain patterns, knots, decay, and irregular growth rings make it difficult for traditional methods to generalize across differ- 1 arXiv:2512.00625v1 [cs.CV] 29 Nov 2025 ent samples. Additionally, certain tree species exhibit non- uniform growth, leading to off-center or distorted pith loca- tions, further complicating automated detection. A major motivation for this study is to identify an opti- mal deep learning approach capable of handling these vari- ations while maintaining high accuracy and generalization. While deep learning has seen widespread success in fields like medical imaging and autonomous navigation, its ap- plication to pith detection remains underexplored. By con- ducting a comparative study of different deep learning mod- els, we aim to determine which approach is best suited for this task. Furthermore, existing research often focuses on either object detection (bounding boxes) or segmentation (pixel- wise classifica

…(Full text truncated)…

This content is AI-processed based on ArXiv data.