Optimised Random Forest Classifier for Fibre Tracking With Fibre Separation in an Orientation Process

ABSTRACT This work presents a fibre tracking process based on a Random Forest classifier. This process is designed to work with minimal labelled data while achieving robust performance. Unlike data‐intensive AI methods, our approach focuses on careful feature selection and iterative optimisation to improve the algorithm. Fibre detection is formulated as a supervised classification task with two classes: fibre and background. Training begins with a small selection of annotated fibres and representative background areas and gradually increases until the trained model shows no significant further changes. Strategic sampling enhances the separation between fibre and non‐fibre areas. This improves stability across images with different lighting and texture. In this work, we optimise the process by evaluating different combinations of features across multiple patch sizes. An average Ground Truth Mask is generated from a combination of models with various features. This significantly improves the detection of challenging fibre types and reduces the number of false negatives compared to our previous implementation. Beyond direct tracking, this approach could become a practical tool for data‐intensive AI algorithms that require large annotated datasets. As this approach produces two classes (fibres and background), we have developed methods to separate individual fibres. We implement a graph‐based separation approach using the skeletonisation of a Random Forest‐generated mask, as well as path scoring based on geometric properties such as length and curvature. This enables us to track the fibres in our experiment using a small amount of annotated data, which contributes to determining their orientations.