Cross-Section Measurement & Distortion Analysis of Multi-Filament Fibers using OpenCV / OpenCvSharp

Python
1

Hi OpenCV Community,

I am developing an industrial desktop application (.NET 8 C# with OpenCvSharp4) for analyzing cross-sectional microscope images of multi-filament synthetic fibers (e.g., Spandex/Elastane).

The goal of the project is twofold:

  1. Filament Counting & Segmentation: Detect, count, and mark each filament centroid in a fiber pack.

  2. Quality & Distortion Metrics (Y-BOT / Laboratory Standard Alignment): Measure individual filament geometry and cross-sectional error across multiple filaments.

I would like to share my current approach, mathematical model, and ask for advice on improving the segmentation accuracy when filaments are tightly packed or partially touching.

1. Mathematical Model & Geometry Pipeline

For each detected filament contour in a cross-section image, we extract two core quality metrics:

A. Distorted ROI (Circularity / Deformation)

Instead of relying solely on standard circularity (4\pi \cdot \text{Area} / \text{Perimeter}^2), we measure width chords along the filament’s principal axis:

  • Long Axis (C): The length of the bounding rotated rectangle (Cv2.MinAreaRect).

  • Short Axis Chords:

    • Short Axis 2 (w\_{50}): Width perpendicular to the Long Axis at the 50\% center point.

    • Short Axis 1 (w\_{25}, w\_{75}): Widths perpendicular to the Long Axis at the 25\% and 75\% length positions.

  • Average Diameter: D\_{\text{avg}} = \frac{\frac{w\_{25} + w\_{75}}{2} + w\_{50}}{2}

  • Distortion Ratio Formula:

    \text{Distorted ROI (\%)} = \left\vert{} 1 - \frac{D\_{\text{avg}}}{C} \right\vert{} \times 100

B. C-Section ROI (Cross-Section Area Ratio / Multi-Filament Error)

For a group of N filaments in a cross-section:

  1. Calculate the pixel area A_i for each filament contour (i = 1 \dots N).

  2. Compute the ratio between the smallest and largest filament area:

    \text{C-Section ROI (\%)} = \frac{\min(A_1, A_2, \dots, A_n)}{\max(A_1, A_2, \dots, A_n)} \times 100

    (Higher percentage indicates better size uniformity among filaments, mapped to quality classes like AA, AB, B based on fiber linear density in Denier/dtex).

2. Current OpenCV Pipeline

Our current image processing workflow (OpenCvSharp4) consists of:

  1. Preprocessing: Downscale to max 1600px \rightarrow Grayscale \rightarrow CLAHE \rightarrow Small Gaussian Blur.

  2. Binarization: Pack-region restricted Adaptive Thresholding (Mean/Gaussian) combined with an Otsu intensity gate.

  3. Morphology: Small 3\times 3 opening followed by max 5\times 5 closing to fill internal voids without merging adjacent filaments.

  4. Distance Transform & Watershed Segmentation:

    • Calculate Distance Transform (Cv2.DistanceTransform).

    • Extract local maxima/peaks to place initial seeds.

    • Apply Watershed segmentation (Cv2.Watershed) to separate touching filaments.

  5. Contour Analysis: Extract contours, fit MinAreaRect, compute axis chords (25\%, 50\%, 75\%), and validate candidates using aspect ratio, circularity, and solidity thresholds.

3. Challenges & Questions for the Community

  1. Over-segmentation / Double-Seed Issue in Watershed:

    In dense areas where filaments have non-uniform internal brightness or irregular shapes (peanut-shaped or flat cross-sections), DistanceTransform often generates multiple peaks inside a single filament, causing Watershed to over-segment (splitting one filament into two markers).

    • What are the best practices or alternative peak-clustering strategies in OpenCV to prevent multiple seeds within peanut-shaped/elongated contours?

  2. Accurate Axis Chord Measurement (%25, %50, %75):

    Currently, we use MinAreaRect to find the principal angle, rotate the contour points, and sweep vertical lines across 25\%, 50\%, and 75\% of the bounding box width to find intersection points with the contour.

    • Is there a more robust/native OpenCV approach (e.g., using cv2.fitEllipse or distance transform ridges) to reliably sample cross-sectional thickness along irregular/non-elliptical contours?

  3. Handling Touching Filament Boundaries:

    When two filaments are physically fused/coalesced, edge gradients become weak. Has anyone achieved stable boundary recovery using pure OpenCV methods before resorting to deep-learning approaches like Cellpose/UNet?

Any insights, code snippets, or suggested OpenCV functions would be greatly appreciated!

Thanks in advance,

Hasan Baş