💻 computer science

Implementation and discussion of the Pith Estimation on Rough Log End Images using Local Fourier Spectrum Analysis method

This paper presents a Python implementation and evaluation of Rudolf Schraml and Andreas Uhl's Local Fourier Spectrum Analysis method for estimating pith locations on rough log end images, validated across two datasets.

Original authors: Henry Marichal, Diego Passarella, Gregory Randall

Published 2026-03-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Henry Marichal, Diego Passarella, Gregory Randall

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a freshly cut log from a tree. If you look at the flat end of that log, you see a series of rings, like a target or a spiderweb. The very center of that web is called the pith.

Why does this matter? For the forestry industry, finding that exact center is crucial. It's like knowing the center of a dartboard before you start playing. If you know where the center is, you can slice the wood perfectly, maximize the value of the lumber, and understand the tree's history.

However, real logs aren't perfect. They have cracks, knots, fungus stains, and the rings aren't always perfect circles. Finding that center point is like trying to find the center of a messy, crumpled spiderweb.

This paper introduces a clever computer program (written in Python) that acts like a super-sleuth to find that center point, even when the wood is messy. Here is how it works, broken down into simple concepts:

1. The "Patchwork Quilt" Strategy

Instead of trying to look at the whole messy log at once, the computer cuts the image into many small squares, like a patchwork quilt.

The Analogy: Imagine you are trying to figure out the direction of a river by looking at a huge map. It's hard to see the flow from high up. But if you zoom in on a tiny square of the map, you can clearly see the water flowing in one specific direction.
The Method: The computer looks at each tiny square individually. Inside each square, the tree rings look like parallel lines.

2. The "Magic Mirror" (Fourier Transform)

Once the computer has a tiny square, it uses a mathematical trick called a Fourier Transform.

The Analogy: Think of this as a magic mirror that turns a picture of "lines" into a picture of "directions." If you have a bunch of parallel lines in the square, this magic mirror shows a bright, glowing line in the opposite direction.
The Result: The computer can now say, "In this tiny square, the rings are running North-South." It does this for every single square on the log.

3. The "Flashlight" Test (Certainty Score)

Not every square is clear. Some have fungus or cracks that confuse the computer.

The Analogy: Imagine the computer is holding a flashlight. If the square is clear, the flashlight beam is bright and steady (High Certainty). If the square is messy, the beam is dim and shaky (Low Certainty).
The Method: The computer ignores the dim, shaky squares and only trusts the bright, clear ones.

4. The "String Theory" (Accumulator Space)

Now, the computer takes all the trusted squares. For each one, it draws a line pointing toward the center (perpendicular to the rings).

The Analogy: Imagine you have a giant map of the forest floor. You take a piece of string for every square you trusted and stretch it across the map, pointing toward where you think the center is.
The Magic: If you stretch hundreds of strings, they will all cross each other somewhere. The place where the most strings cross is the pith. It's like a "crowd vote" where the center of the crowd is the answer.

5. The "Fuzzy Search" (Peak Finding)

Finally, the computer looks at the map to see where the most strings crossed.

The Analogy: Imagine a pile of sand where the strings crossed. The highest pile of sand is the answer. The computer smooths out the pile a little bit (like a gentle breeze) to make sure it finds the absolute highest peak, not just a small bump. That peak is the estimated center of the tree.

How Well Does It Work?

The authors tested this "super-sleuth" against other methods using two different sets of wood photos:

The "Messy" Set (Uruguay): These logs had fungus, cracks, and bad lighting.
- Result: The new method was very good at finding the general area of the center, even if it wasn't pixel-perfect. It was more consistent than the other methods, which sometimes got lost and pointed to the wrong ring entirely.
The "Clean" Set (Germany): These logs were very clean and clear.
- Result: Both methods worked almost perfectly.

The Bottom Line

This paper is about building a tool that helps computers "see" the heart of a tree, even when the tree is old, cracked, or dirty.

Why it's cool: It's fast (less than a second per log) and doesn't need expensive supercomputers; it runs on a regular laptop.
The Catch: If the tree has a huge black fungus stain right over the center, the computer might get confused, just like a human would. But for most logs, it's a very reliable way to find the "bullseye."

In short, this method turns a messy, complicated image of a tree trunk into a simple math problem, allowing us to find the center of the tree with the precision of a laser, but the logic of a detective.

1. Problem Statement

The accurate detection of the pith (the geometric center of a tree's cross-section) is critical for the forestry industry, particularly for industrial log manipulation and automated tree-ring delineation.

The Challenge: Many ring-delineation algorithms rely on a "spider web" model, assuming rings are concentric circles intersecting at the pith. However, real-world wood slices often contain asymmetries, cracks, knots, and fungal stains (blue stain), which disrupt this ideal pattern.
The Goal: Develop a robust algorithm capable of estimating the pith location despite these perturbations, without relying on deep learning models that require massive training datasets.

2. Methodology

The paper implements and analyzes the method proposed by Schraml and Uhl (2013), which utilizes Local Fourier Spectrum Analysis. The pipeline consists of four main stages:

A. Pipeline Overview

Input: An RGB image of a wood cross-section and a background mask.
Local Orientation Estimation: The image is divided into overlapping patches (blocks).
Accumulator Space Construction: Lines representing ring normals are traced across the image to find intersections.
Peak Detection: The pith is identified as the global maximum (or average of maxima) in the accumulator space.

B. Local Orientation Estimation (The Core Innovation)

Instead of using gradient-based methods (like HOG), this method leverages the 2D Fourier Transform:

Hypothesis: In a local patch, tree rings appear as roughly parallel curves. The Fourier Transform of such a region concentrates energy along a line perpendicular to the direction of the rings (the normal direction).
Process:
1. Splitting: The image is split into blocks (e.g., 100x100 pixels) with a defined overlap (e.g., 20%).
2. Fourier Analysis: The 2D FFT is computed for each patch.
3. Filtering: A band-pass filter removes low frequencies (noise/background) and high frequencies. A threshold filters out weak spectral peaks.
4. Line Estimation: The dominant orientation in the frequency domain is estimated using one of four methods:
  - Peak: Connects the highest magnitude frequency to the center.
  - LSR (Least Squares Regression): Fits a line to the spectrum.
  - WLSR (Weighted LSR): Weights the fit by the square root of magnitude.
  - PCA (Principal Component Analysis): Extracts the principal axis.
5. Certainty Scoring: Each estimation is assigned a confidence score (e.g., $|R^2|$ for regression, eigenvalue ratio for PCA). Only lines exceeding a certainty threshold are kept.

C. Accumulator Space & Peak Finding

Accumulation: The estimated normals are traced across the entire image. An accumulator matrix counts how many lines intersect or pass through each pixel.
- Intersection Logic: Increments pixels where two lines cross.
- Pass-through Logic: Increments pixels where a line passes.
Peak Detection: A Gaussian filter is applied to smooth the accumulator space. The pith location is determined by the global maximum (or the mean of multiple global maxima).

3. Key Contributions

Python Implementation: The authors provide a complete, open-source CPU-based Python 3.11 implementation of the Schraml & Uhl (2013) algorithm, making it accessible for reproducibility and further research.
Comprehensive Evaluation: The method was rigorously tested on two distinct datasets:
- UruDendro: 64 images of Pinus taeda with significant challenges (fungus, cracks, varying lighting).
- Kennel: 7 images of Abies alba.
Parameter Optimization: The study performed a grid search to identify optimal hyperparameters (patch size, overlap, orientation method, thresholds) for different dataset characteristics.
Comparative Analysis: The method was benchmarked against the state-of-the-art Ant Colony Optimization (ACO) method (Decelle et al.) and discussed in the context of Deep Learning approaches.

4. Results

The performance was evaluated using Euclidean distance (pixels), Normalized Euclidean distance, and Precision (True Positive rate within the medulla).

Metric	UruDendro Dataset (Challenging)	Kennel Dataset (Cleaner)
Mean Distance Error	Schraml (44px) vs ACO (72px)	Schraml (8px) vs ACO (6px)
Precision (TP within Medulla)	ACO (70%) vs Schraml (14%)	Schraml (71%) vs ACO (71%)
Robustness to Noise	Schraml failed significantly on images with heavy fungus/cracks near the pith.	Both methods performed well.
Execution Time	< 1 second per image (CPU).	< 1 second per image (CPU).

Key Findings:

On Clean Data: The Schraml method performs comparably to or better than ACO in terms of raw distance error.
On Noisy Data (UruDendro): While Schraml has a lower mean error, it suffers from a high False Positive rate (14% precision) when fungus or cracks obscure the pith. The ACO method is more robust in these specific "failure" cases, keeping detections within the medulla more often.
Failure Mode: The Schraml method tends to fail when the ring structure near the center is completely destroyed (e.g., black fungus covering the pith), as the local Fourier spectrum cannot establish a valid orientation.

5. Significance and Conclusion

State-of-the-Art Relevance: Despite being published in 2013, the Schraml & Uhl method remains competitive with modern techniques, offering a lightweight alternative to Deep Neural Networks (DNNs) which require extensive training data and GPU resources.
Efficiency: The method runs in under 1 second on a standard CPU, making it suitable for real-time industrial applications.
Limitations: The method is highly dependent on the visibility of ring structures near the center. Heavy fungal infection or physical damage at the pith location significantly degrades performance.
Future Work: The authors suggest that while the method is robust to general perturbations, hybrid approaches or specific preprocessing for fungal stains might be necessary to improve precision in the most difficult forestry scenarios.

In summary, this paper validates the Local Fourier Spectrum Analysis approach as a fast, effective, and mathematically elegant solution for pith detection, provided the wood surface retains sufficient ring structure near the center.