Automatic Wood Pith Detector: Local Orientation Estimation and Robust Accumulation

TL;DR

This work tackles robust automatic pith localization in wood cross‑sections across gymnosperm and angiosperm species under diverse imaging conditions. It introduces three complementary approaches: APD, a structure‑tensor and optimization‑driven method; APD‑PCL, which uses PClines and RANSAC to handle degraded ring patterns; and APD‑DL, a YOLOv8 detector trained with five‑fold cross‑validation on diverse datasets. The authors release the UruDendro2/UruDendro3 datasets and demonstrate state‑of‑the‑art performance with real‑time capability on CPU for the classical methods, while APD‑DL achieves strong results on GPU. The work advances robust pith detection across species and perturbations, with practical impact for tree age estimation and wood processing, and provides supplementary materials detailing parameters and difficult‑case analyses. Key equations include the structure tensor based local orientation $ST_O$, coherence $ST_C$, and the pith optimization $c_{opt}$, which collectively enable precise pith localization under challenging conditions.

Abstract

A fully automated technique for wood pith detection (APD), relying on the concentric shape of the structure of wood ring slices, is introduced. The method estimates the ring's local orientations using the 2D structure tensor and finds the pith position, optimizing a cost function designed for this problem. We also present a variant (APD-PCL), using the parallel coordinates space, that enhances the method's effectiveness when there are no clear tree ring patterns. Furthermore, refining previous work by Kurdthongmee, a YoloV8 net is trained for pith detection, producing a deep learning-based approach to the same problem (APD-DL). All methods were tested on seven datasets, including images captured under diverse conditions (controlled laboratory settings, sawmill, and forest) and featuring various tree species (Pinus taeda, Douglas fir, Abies alba, and Gleditsia triacanthos). All proposed approaches outperform existing state-of-the-art methods and can be used in CPU-based real-time applications. Additionally, we provide a novel dataset comprising images of gymnosperm and angiosperm species. Dataset and source code are available at http://github.com/hmarichal93/apd.

Figures (8)

• Figure 1: (a) to (c) Some examples from UruDendro datasetUruDendro, (d) The whole structure, called spider web, is formed by a center (the slice pith), rays, and the rings (concentric curves). In the scheme, the rings are circles, but in practice, they can be (strongly) deformed as long as they don't intersect another ring.
• Figure 2: Cost function definitions
• Figure 3: Principal steps of APD method (L04d image from UruDendro2 collection). (a) Resized slice image, without background; (b) Sampled LO produced by the Structure Tensor estimation; (c) Accumulation space defined by the LO supported lines; (d) Plot of the cost function (\ref{['eq:cos2']}), highest values in yellow; (e) Sub image built around the solution $c_1$ obtained after the first iteration; (f) Evolution of $c_i$. The final solution is in blue; previous iterations' solutions are in red.
• Figure 4: Use of PClines to cluster converging local orientations for slice F07e (same as \ref{['fig:F07e']}). (a) Local orientations; (b) Selection of the converging segments in the twisted space using RANSAC to fit a line (in red). Inliers are colored in green; (c) The same procedure is applied in the straight space; (d) In blue, the converging LO (inliers from both sub-spaces) and the LO to be removed in red.
• Figure 5: LO Accumulation space and cost function for slice F07e with and without applying the PClines filtering method. (a) LO Accumulation space with no filtering; (b) cost function of LO with no filtering; (c) LO Accumulation Space with PClines filtering; (d) cost function of LO with PClines filtering