Insegt Fibre: a user-friendly software for individual fibre segmentation

Insegt Fibre is a software toolbox for volumetric fibre segmentation. The toolbox comes with scripts to detect the centres of individual fibres in 2D and 3D from tomograms acquired through X-ray imaging, and a graphical user interface to verify the accuracy of the resulting 3D tracks. In addition, there is a script to characterise fibre orientations in 3D and a script to match corresponding fibres across a 4D time-lapse sequence, which enables the characterisation of composite micro-structural changes of individual fibres. Insegt Fibre is based on a segmentation method that uses a dictionary of image patches which has been trained to model the patterns/features that are repeated in the image data at a certain scale defined by the patch size. Thus, the dictionary-learning segmentation algorithm has proven highly successful in modelling fibres that have regular cross-sections, commonly found in fibre reinforced composite materials. The algorithm is robust to noise and artefacts in the data and therefore excels in measuring fibre geometry from a variety of scan qualities, even when fibres are densely packed and the boundaries between them are unclear. Insegt Fibre is simple to use, it comes with a manual and it requires minimal input from the user. Besides presenting the recent user-friendly version of this robust method for measuring the 3D tracks of fibres from X-ray CT data, the paper gives an overview of the possibilities that the method gives with regards to characterisation of composite micro-structure and fibre behaviour under load. Thanks to the precision to which fibre geometry can be characterised with this method, it is now possible to follow how each individual fibre changes across data-sets acquired under progressive loading conditions. All in all, Insegt Fibre makes image-based characterisation of fibrous materials simpler, more accessible and applicable to a broader range of studies. 1 MOTIVATION AND SIGNIFICANCE Fibre-reinforced composites are employed in a wide range of technologies for their high strength and stiffness relative to their weight. As the strength and stiffness properties of a composite are determined by its internal three-dimensional (3D) structure, X-ray computed tomography (CT) is becoming a popular tool for composite characterisation [1]. The recent advances in X-ray micro-CT have led to an increase in image quality for a given acquisition time and sample size or field of view (FoV). In particular, the improvement in spatial resolution has enabled the observation of individual fibres in scans capturing a larger FoV – large enough to be representative of the fibre structure. If interested in discovering biases in the process of fibrebundle manufacturing, a representative FoV would contain a full bundle with tens of thousands of fibres. Composite behaviour under load can be studied in-situ both at laboratory scanners and synchrotron beamlines. It is done by acquiring images while a sample is loaded, which makes it possible to monitor the geometric changes in the composite. Such data is 4D with three spatial and one temporal dimension. Thanks to the increase in temporal resolution obtained at the ultra-fast imaging set-ups found at some synchrotron beamlines, even the very small but sudden microstructural changes that lead to composite damage can be captured. This includes the moments in which the damage mechanism actually initiates, as well as the moments just before the sudden failure of the composite. Hence, the analysis of these 4D time-lapse data-sets will accelerate the understanding of the very complex damage mechanisms that Monica J. Emerson, Anders B. Dahl, Knut Conradsen and Vedrana A. Dahl affect fibre composites. To measure the structural details of the vast amount of fibres contained in a bundle, or the very small changes across load steps in 4D data, quantitative image analysis is required. The first step is typically the detection of individual fibres using image segmentation. The accuracy of this step is crucial, as it defines the geometry of the material phases on which the subsequent analyses will be based, and ultimately the material characterisation. Several methods have been proposed for detecting fibres in scans acquired with synchrotron light, e.g. in [2-3]. In both studies, individual fibres are segmented by setting a threshold on the intensities of the image. A segmentation based on intensity thresholding is however highly susceptible to noise and artefacts in the image, as well as to the contrast between the material phases. For composites where fibres are densely packed, more robust image segmentation methods are necessary to handle a variety of image qualities. Methods that are relatively robust to noise and artefacts will facilitate the analysis of e.g. fast scans acquired while loading a composite in-situ, or scans acquired with relatively coarse pixel sizes (4 px/fibre) to image a larger (and thus more representative) sample region in a single scan. What is more, lower requirements to the scan quality translate into lower requirements for the X-ray set-up and shorter scan times. In other words, robust image segmentation algorithms will both broaden the application of composite imaging and make it accessible to a larger community of scientists. More elaborate fibre segmentation algorithms have been proposed for lower quality scans of densely packed fibres in [4-6]. While Czabaj et al. [4] and Sencu et al. [5] developed complex tracking methods, our method described in Emerson et al. [6], has until now focused on the robustness, automation and speed of the fibre detection step. The dictionary-based segmentation method is able to analyse images of low resolution and with high noise levels where a good accuracy is obtained for a large range of parameter values, as demonstrated in [7]. In the template matching segmentation approach used by Czabaj, six fibre cross-sections are selected from the data manually to represent the appearance of the fibres. Instead, the probabilistic dictionary segmentation approach used here learns the representative patches from the data. The number of patches employed to model the data-set is much larger (in the order of hundreds) and can be increased to model a larger variety of fibre cross-sections. For example, depending on the local inclination of the fibres with respect to the scanning axes, the fibres may vary in size (diameter) and shape (e.g. ellipse axes ratio). This paper presents Insegt Fibre, an interactive software tool for detection of individual fibres in 3D, that is now available online for free. The software is based on the latest version of the dictionary-based segmentation algorithm, first presented for 3D detection of individual fibres trajectories in [6]. 2 SOFTWARE DESCRIPTION Insegt Fibre is a software toolbox to measure individual fibres from X-ray CT scans of fibre-rich composites. Section 2.1. gives an overview of the methodology and algorithms on which the toolbox is based. Section 2.2 describes the input that is necessary from the user. To end with, Section 2.3 details the material that comes inside the Insegt Fibre toolbox. 2.1 Fibre detection algorithm The aim of this section is to describe the changes in methodology since [6]. To understand the changes, a short overview of the method is provided, and the relevant dictionary concepts are presented. For more information, the reader is referred to [6]. 2.1.1 The pipeline for measuring the 3D tracks of individual fibres Insegt Fibre is based on a method that detects the 3D centre lines of individual fibres in two steps, these are shown in Fig. 1. The first step concerns the detection of fibre cross-sections in 2D images. In the case of unidirectional fibre-reinforced composites, the 3D scan can be sliced such that the resulting 2D images show fibre cross-sections of similar shape (circular) and size (diameter). The second step concerns tracking, i.e. matching centre points from one slice to the next, so as to define determine the 3D centre lines of the fibres contained in the composite volume. Figure 1: The Insegt Fibre pipeline. In the first step, fibre cross-sections are detected from 2D images. This process involves i) slicing the 3D scan in the direction orthogonal to the fibres and ii) applying the dictionary model learnt from the data to find the centres of the fibre cross-sections at each 2D slice. In the second step, the 3D fibre tracks are defined by connecting the fibre cross-sections that correspond to the same fibre. This step involves matching fibre detections along the third dimension, based on the cross-sectional distance between the points to be matched. 2.1.2 Detection of fibre cross-sections from 2D images The first step concerns finding the coordinates of the fibre cross-section centres at every 2D slice. Fig. 2 illustrates this process. A dictionary model learnt from a training data-set is applied to an intensity image (see Fig. 2(a)) to obtain its probabilistic segmentation (see Fig. 2(b)). The probabilistic segmentation (or probability map) indicates how likely it is that a pixel belongs to the central region of a fibre. The segmentation of fibre-centre regions is obtained from the probability map by setting a threshold on the probability value. That is, all the pixels with a probability above the threshold value are labelled as belonging to the fibre-centre class. The fibre-centre class (pink in Fig. 2(c)) grants the segmentation of the individual fibres. The coordinates of the fibre centres (black dots in Fig. 1(c-d)) are computed as the centroids of the pink connected components. The training data-set comprises a 2D image that has been annotated by the user, to indicate whether a pixel belongs to a fibre-centre region (i.e. is close to a fibre centre) or not. The dictionary of images patches is comprised of a dictionary of intensity patches, and a corresponding dictionary of label patches. In the fibre case, the labels indica