A Brief Review of Hemp Fiber Length Measurement Techniques

Abstract

Accurate fiber length measurement is essential for the processing and quality management of textile products. This article reviews the current methods used to measure fiber length, including manual, photoelectric, capacitive, and optical techniques. Existing sample preparation processes for natural fiber characterization have been primarily developed for cotton and wool fibers. However, hemp fibers present unique challenges due to their greater length variability, high strength, and low elongation, making some traditional sample preparation methods less effective. Image processing offers a promising approach for scalable and precise measurement of hemp fiber length. Nevertheless, current image processing techniques are limited by the inability to effectively handle overlapping fibers, which increases both the time and cost of testing. Continued research into developing more advanced segmentation algorithms could lead to more widely adopted commercial methods for fiber measurement.

1. Introduction

Fiber length measurement is a crucial parameter in the characterization of natural fibers, as it plays a vital role in determining the parameters for yarn formation. The length of the longest fiber sets the distance between the draft rollers, which draw fibers into sliver and yarn. Incorrect machine settings can result in fiber breakage or cause variations in yarn weight, making fiber length distribution a key quality factor for both the processing stage and the final yarn quality [1]. Longer and finer fibers contribute to greater yarn evenness and strength, and producing thinner yarns relies on the presence of longer fibers. Traditionally, most methods for measuring the length and fineness of fiber samples have been designed for cotton or wool. These fibers generally have a narrower length distribution and are shorter than other natural fibers, such as linen or hemp.

The hemp plant, Cannabis sativa, has gained increasing popularity, as evidenced by a global cannabidiol market size of USD 7.71 billion in 2023 and a projected compound annual growth rate of 15.8% from 2024 to 2030 [2]. In the textile industry, hemp is becoming a preferred sustainable fiber alternative due to its short growth cycle and carbon-sequestering properties. The cultivation of hemp is also more cost-effective, requiring 77.63% less expenditure than cotton, primarily because it needs significantly less fertilization and irrigation [3]. Hemp fibers share similarities with flax and other bast fibers, characterized by high tenacity and low elongation, and are valued for their versatility in a wide range of applications, including textiles, technical uses, paper, oils, and composite materials. Specifically, hemp composite, yarn, and textile applications will rely on accurate fiber length measurements to determine mechanical performance both during processing and for the final product [3,4,5]. However, due to the prohibition of hemp cultivation in many countries throughout the 20th century, there has been a lack of standardized methods for characterizing these fibers. Moreover, hemp fibers, with their long length, variability, and brittleness, are less suited to the characterization and processing methods developed for cotton or wool [5,6].

There is a clear need to evaluate current fiber length measurement methods to determine their suitability for hemp fibers. This review provides an in-depth analysis of existing standards and research on fiber length measurement, with a focus on their application to hemp fibers. It also summarizes alternative methodologies, such as image processing techniques, that could be adapted for hemp fiber measurement. Finally, the review discusses future research directions for advancing fiber length measurement of hemp.

The fiber length distribution illustrates the variation in fiber lengths within a sample, typically presented in a graphical form. This distribution can be calculated based on either weight or number. For a weight–length distribution, fibers are sorted into groups of similar lengths, and each group is weighed. The weight fraction of each length group is then plotted against the fiber length. In contrast, a number-length distribution is created by counting the number of fibers in each length group or by multiplying the weight fraction by the individual fiber weight. The resulting number fraction is then plotted against the fiber length. As demonstrated in Figure 1 below, plotting both the number-based and weight-based distributions for the same sample reveals some differences. The number-based distribution shows a higher frequency of short fibers, but since short fibers contribute little to the overall weight, their representation is much lower in the weight-based distribution. These distributions provide valuable insights into the uniformity and maximum and minimum fiber lengths within the sample.

Key fiber length parameters include mean length, span length, maximum fiber length, upper quartile length, short fiber content (SFC), and uniformity index (UI). The mean length represents the average fiber length and can be calculated based on either number or weight. Span length refers to the distance covered by a certain percentage of fibers; for example, the 2.5% span length indicates that 2.5% of the fibers in the sample are of the specified length or longer and is often used to determine the maximum fiber length. The upper quartile length is the average length of the longest quarter, or 25%, of fibers. Short fiber content (SFC) is the percentage of fibers measuring less than 0.5 inches and can also be measured by weight or number. The uniformity index (UI) is the ratio of the mean length to the upper half mean length; a higher UI indicates more uniform fibers. A fibrogram, as shown in Figure 2, displays the cumulative percentage of fibers that are equal to or shorter than a given length, allowing for a quick visual assessment of span lengths and mean lengths.

2. Methods of Fiber Measurements

2.1. Manual

The primary method for manually measuring fiber lengths involves brushing and individualizing a bundle of fibers using parallel combs, which help align the fibers for sorting by length. Once parallelized, the fibers are grouped by size, and each group is measured by number or weight to determine the weight fraction. When a device is added to assist in sorting, this is known as the array method, which uses a comb sorter—the first objective tool developed for fiber length measurement. Popular methods that use the array method include the Baer, Suter–Webb, and Shirley comb sorters, which are considered among the most accurate for measuring fiber lengths.

In the array method, a set weight of fibers, typically 75 mg, is sorted on a velvet board in increments of 0.125 inches. Parallel combs are used to align the fibers, which are then manually sorted into length groups from longest to shortest using tweezers. Each group is weighed to determine the length–weight distribution, which can then be used to calculate other measurements, such as mean fiber length, upper quartile length, and short fiber content.

The WIRA fiber machine (Wira Instrumentation, Bradford, UK) was developed to measure wool fibers individually rather than in groups. Initially, the fibers are manually individualized with combs. A fiber is then drawn through the machine, and its end is placed on a rotating shaft with a spiral groove, as illustrated in Figure 3. As the fiber moves through two pressure plates, steady tension is maintained. A wire resting on top of the fiber drops and completes an electrical circuit when the end of the fiber passes under it, causing the rotating shaft to stop. Tweezers lift the fiber to a mechanical counter above, arranged in 0.5 cm sections, to record its length. These lengths are accumulated, and the results are presented as a length distribution curve and mean length. This semi-automatic system can measure up to 500 fibers per hour [7].

Manual array methods are highly time-consuming, taking 2–3 h per sample, and are therefore costly to perform. The combs used in the Baer Sorter, for example, are designed to accommodate specific fiber lengths, typically those of wool or cotton, and may not be suitable for measuring longer bast fibers like hemp. While these methods provide direct measurements, their accuracy—especially for shorter fibers—largely depends on the operator’s skill. Consequently, while manual array methods could be used with hemp fibers, they are not ideal for commercial applications where time and cost efficiency are critical.

2.2. Photoelectric

Two common photoelectric fiber length measurement tools are the Uster Fibrograph (Uster, Switzerland) and the Shirley photoelectric stapler. In both methods, a fiber beard is created by randomly clamping a bundle of vertically aligned fibers from a roving or sliver sample using a fibrosampler. The fibers are then scanned, and light intensity is measured at each point to determine the density of fibers at different lengths. The amount of light blocked is proportional to the number of fibers in that area and is recorded graphically as a fibrogram. The resulting data are typically analyzed to provide percentages of span length and a uniformity index.

The High Volume Instrumentation (HVI, Uster Technologies, Uster, Switzerland) uses similar technology to the Fibrograph but includes additional testing features, such as fiber strength, elongation, and color. It is currently the USDA standard method for testing cotton bale samples due to its fast testing rate [8].

However, this method has limitations. Since longer fibers have a higher chance of being clamped during beard preparation, the measurement can be length-biased. Fiber crimp and the clamping length can also reduce measurement precision [9]. Additionally, this method assumes uniform fiber widths, which is generally valid for more consistent fibers like cotton but may not be as accurate for others. The requirement for a fiber beard makes this method less suitable for hemp fiber measurement. Hemp fibers are longer and more brittle than cotton fibers, making them prone to breakage during this process, which could reduce accuracy. Furthermore, any method that uses a clamping device is prone to introducing some measurement errors.

2.3. Capacitive

The Peyer Almeter is a capacitive method for measuring fiber length, originally developed for wool. A Fibroliner (Uster Technologies, Uster, Switzerland) is used to prepare the sample by mechanically combing the fibers to align and parallelize them at one end. The fibers are then placed between plates for measurement. The Almeter scans the fiber beard every 0.24 mm using capacitive measurements, where the output is proportional to the mass distribution, assuming a constant linear density [10]. This analysis process takes approximately 15 min per sample. The Wira Fiber Diagram (Wira Instrumentation, Bradford, UK) operates on the same principle as the Peyer Almeter, and both have standardized testing procedures under the International Wool Textile Organization (IWTO).

Similar to photoelectric methods, the short fiber content measured with this device tends to be lower than the actual percentage. This discrepancy is likely due to the sample preparation process, where shorter fibers may not align at the bottom, migrate toward the center, or be lost during combing [11]. Capacitance-based measurements are also prone to errors caused by moisture content or foreign particles. Since both methods depend on parallelizing the fibers and creating a fiber beard, they are not suitable for measuring the length of hemp fibers.

2.4. Optical

The Advanced Fiber Information System (AFIS, Uster Technologies, Uster, Switzerland) was developed to address the limitations of existing automatic fiber length measurement systems in accurately measuring short fiber content (SFC). To achieve more precise measurements of SFC, the AFIS system includes an opening action that individualizes and aligns the fibers, which are then transported by airflow into the sensor area. A light beam is directed through the sensor area, and as fibers pass through, they scatter and block the light, as illustrated in Figure 4. The length of each fiber is determined by measuring the duration that the fiber blocks the light, based on the assumption that all fibers are moving at the same velocity [12].

There are several instances where data from the AFIS have been shown to contain errors. For example, the carding action used in the AFIS can cause fiber breakages, artificially increasing the short fiber content (SFC). Additionally, any entanglements or crimps in the fibers can result in measurements that are either too long or too short. This method is not ideal for hemp fiber testing, as the opening action is likely to cause fiber breakages, leading to inaccurate measurements.

The Fibrotest (TexTechno, Mönchengladbach, Germany) measures both fiber length and fiber strength consecutively on the same sample. This process is semi-automated: the operator first prepares a sample on a holder using a comb and then places it into the Fibrotest. Alternatively, a Fibrosampler can be used to create the fiber sample beard. The Fibrotest employs an optical sensor with a laser and camera to scan the fiber bundle longitudinally as illustrated in Figure 5, with a scanning width of 60 mm and a maximum travel distance of 190 mm [13]. After the sample is weighed and the length distribution is determined, other fiber length measurements can be calculated [14]. However, as with other methods, the sample preparation process used here is likely unsuitable for hemp fibers.

2.5. Image Analysis/Processing

In his 2000 PhD thesis, Ikiz developed an image analysis software to measure fiber lengths, examining the effects of sample preparation, lighting, resolution, preprocessing, and processing. To prepare samples, Ikiz photographed fully individualized fibers and some fibers with partial overlaps using a CCD camera, ensuring even light distribution during the setup [9].

The images were processed by converting them to binary using local thresholding. Outline detection, thinning, and addition algorithms were then applied. These steps could introduce noise and create short branches in the images, leading to errors, especially in cases of fiber crossover. Thinning could also result in “broken skeletons,” where pixels along the fiber length are removed due to the thinning or thresholding processes. To address this, additional steps of dilation and erosion were used to reconnect broken skeletons. However, if fibers were too close together in the image, they could be incorrectly connected during this process, resulting in further errors.

To calculate the final fiber length, Ikiz employed a chain code algorithm to identify and count the number of adjacent pixels corresponding to each fiber. He addressed crossovers in the fiber images by examining all white pixels with more than two connections after the thinning process. These connection points were assigned a gray value, and the software iterated over them to count and match connection points; any areas with unmatched crossovers were deleted.

Ikiz found that the thinning and addition algorithm provided the highest accuracy and precision for images without crossovers. However, the outline algorithm with a 4-neighborhood threshold was the fastest and still met his precision limit of 0.5 mm. Ikiz also compared different image resolutions and lighting conditions, finding that front lighting produced more accurate results for images without crossovers, while backlighting was the most reliable for achieving accuracy across all variables [9].

In 2008, Wang et al. built on Ikiz’s work by developing a new prototype that uses a scanner to measure cotton fiber length [15]. In their process, fibers are manually individualized and placed on a glass slide for scanning to capture images. They used both images with completely separated fibers and those with higher fiber content that included crossovers. The first step in their process is image binarization. A 3 × 3 neighboring filter is applied to reduce noise in the background and create more uniform fiber intensity. Given that the scanned images have consistent lighting, a global thresholding filter is used to binarize the image. Morphological closing is then applied to restore any broken edges of the fibers caused by binarization. Next, the medial axis of each fiber is determined, resulting in a skeletonized image where the fibers are reduced to a single pixel width by progressively removing boundary pixels, as illustrated in Figure 6. The fibers along these axes are then categorized by numbering adjacent pixels to identify points where inter− or intrafiber crossovers occur. The connections at these crossover points are deleted, and the junction points are labeled. Finally, these junction points are restored by reconnecting the pixels, allowing for the determination of which fiber segments belong to the same fiber [15].

To reduce overestimation when measuring fiber lengths, Wang et al. employed a piecewise polynomial fitting strategy [15]. This approach minimizes the effects of minor crimps and directional changes in fibers that could artificially increase the measured fiber length. First, critical points where the fiber changes direction are identified, and polynomials are fitted to the fiber’s path. The lengths of these fitted polynomial segments are then summed to determine the total fiber length.

Their results showed that the relative error percentage decreased using their proposed algorithm, as verified by measuring filament fibers cut to predetermined lengths. They also compared this method to the AFIS by measuring cotton fibers in both systems. The mean length obtained by their method was longer, and the short fiber content (SFC) was lower, consistent with findings that AFIS tends to underestimate fiber length, leading to a global underestimation of length and overestimation of SFC. Wang et al. found that their system significantly improved upon Ikiz’s method for measuring individualized fiber lengths [15].

In 2020, Hong and Wang introduced a new technique to identify fiber crossovers in images [16]. They used a microscope to capture images of wool fibers, which were then pre-processed by converting them to binary and applying thresholding segmentation and denoising to extract the fibers. After identifying the crossover points, they applied a local binary pattern algorithm to the original grayscale image, creating a binary pattern for each pixel based on the luminosity of its surrounding 8 pixels. Assuming that each fiber would have a more consistent local binary pattern than others, they compared the values on each side of the crossover points. Using the Hamming distance to compare these values, they matched similar values to determine which segments belonged to each fiber. This is illustrated in Figure 7, showing an example of what a fiber image would appear before and after identification of fiber segments. Hong and Wang successfully aligned fiber crossovers and measured fiber diameters in different batches of wool [16].

The Fibreshape (IST AG, Ebnat-Kappel, Switzerland) is currently the only commercialized machine that uses image processing to measure fiber lengths. It measures both length and fineness and is available in two options: one requiring manual sample preparation and another with an automatic feeder. Both systems can measure fibers ranging from 2 µm to 30 cm in length [17]. The fibers are scanned, and an image analysis algorithm determines their length, followed by a statistical evaluation of the data. During processing, the system identifies fiber crossovers using eroding and neighborhood operations, deletes the crossovers, and uses dilation to reconstruct the fibers, measuring their length and width. Müssig and Amaducci successfully measured hemp fiber fineness using the Fibreshape machine [18]. While this machine can likely measure a variety of fiber types, it lacks support for identifying fiber crossovers and is limited to measuring fiber lengths up to 30 cm.

These more manual image processing methods show promise for hemp fiber applications. However, as the authors noted, most algorithms currently lack solutions for handling multiple fibers occupying the same space, fibers that touch but do not overlap, fibers that overlap along a large portion of their length, and multiple fiber crossovers. Identifying crossovers and loops remains the most significant challenge in image processing. While Hong and Wang’s method is the most recent, it did not address whether it could correctly identify looped fibers. Although such situations may not occur in the very sparse images used in their research, they would pose issues if many fibers were measured in a single image. To accurately sample a large batch of fibers, a sizable random sample must be tested. In their images, the fiber density is very low, meaning that numerous images would need to be analyzed to provide a reliable estimation of the overall fiber length. Further research is needed to improve the robust identification and classification of fiber loops and crossovers and to increase the density of fibers analyzed in a single image.

Another recent trend in measuring fiber length distributions using image processing involves the use of cotton dual beard fibrographs. In 2020, Zhou et al. proposed a method that utilizes the Levenberg–Marquardt algorithm [19]. The dual-beard sample can be prepared using the Lengthcontrol system (LCT, TexTechno, Mönchengladbach, Germany), which automatically creates samples from a sliver clamped at the center with tapered edges on both sides, illustrated in Figure 8a. This sample is then scanned and analyzed to determine length distribution, upper length, short fiber content, and other parameters. Once the dual-beard sample is prepared, it is scanned to obtain a high-resolution image with consistent lighting. The clamping line is designated as the 0th column, and each column of pixels is analyzed to determine the average grayscale value. These grayscale values are then normalized by dividing each by the maximum subtotal across all columns. Each column now represents the relative number of fibers at each length, given the known pixel-to-millimeter conversion. The fibrograph is plotted with length on the x-axis and relative intensity on the y-axis as shown in Figure 8b, illustrating the proportion of fibers on both sides of the beard. Thus, each point on the graph indicates the relative number of fibers that have a length greater than the distance from the centerline.

Zhou theorized that, in an ideal scenario where fibers of uniform length are evenly distributed and randomly clamped perpendicularly, the resulting dual beard would form the shape of a parallelogram. The corresponding fibrograph of this parallelogram would appear as an isosceles triangle, where the baseline equals twice the fiber length, and the height matches that of the parallelogram, as illustrated in Figure 9. Thus, the dual beard can be conceptualized as a group of parallelograms, and the fibrograph can be approximated as a series of isosceles triangles. By estimating a series of triangular base functions that best fit the original fibrograph function, the fiber length distribution can be more accurately determined. Given that this is a non-linear optimization problem, the authors selected the Levenberg–Marquardt algorithm to achieve precise results [19].

Zhou et al. verified the accuracy of their method by creating samples of cut cotton fibers that were combed and evenly distributed into parallelogram shapes. The parallelograms were designed to have the same height, allowing the assumption that each contained roughly the same number of fibers. When these images were analyzed, the resulting fibrographs formed the shape of an isosceles triangle with a base length twice that of the fiber length. While the graphs showed minor peaks corresponding to other fiber lengths, the authors attributed these to human error in cutting the samples. Overall, the graphs accurately estimated the lengths of the cut fibers. They then tested their method using slivers with and without short fibers. The resulting fiber length distributions accurately reflected changes in the short fiber content (SFC). Finally, they compared their results with those obtained using the AFIS system by testing three cotton samples with both methods. They observed a relatively strong correlation between the two systems, although there were noticeable differences in the fiber length distribution graphs. The AFIS showed fibers up to 60 mm, while the dual beard fibrograph (DBF) showed fibers up to 38 mm. This discrepancy is likely due to a known issue with the AFIS, where fibers may not fully separate during the opening process and are measured as a single, longer fiber. Additionally, the AFIS reported a higher SFC, which is also attributed to increased fiber breakage during the opening action [20].

In 2020, Lang et al. further refined this method by applying it to measure both cotton and wool fiber beard samples, introducing a new algorithm to convert relative linear density curves into length distributions [21]. This new algorithm avoids the use of differential operations, which can introduce errors in the curve. When comparing their results with manually measured length distributions by weight, the differences were minimal—0.6% for cotton and 0.7% for wool [21]. In 2021, Lang et al. used the same process with another new algorithm to determine the weight-based short fiber contents (SFC) of raw cotton and cotton slivers [22]. Their results demonstrated accurate SFC measurements with just two to four specimens per sample [22].

Overall, dual beard fibrography is best suited for cotton and wool fibers. Due to its similarities in fiber preparation with other fiber beard measurement systems, it may introduce errors when applied to hemp fibers. Additionally, the assumptions of uniform fiber weight used in this method may not be accurate for hemp fibers, which tend to be more irregular than other fibers.

2.6. Machine Learning and Deep Learning

Recently, 3D image analysis has also been employed to characterize fibers, particularly in non-woven or composite materials where the fibers are very small and thin. In a 2022 study, Zanini and Carmignato used X-ray computed tomography (CT) scans to determine fiber length. After acquiring images using a microCT scanner, they reconstructed a digital model of the fibers using CTPro 3D software 3.1.9, (Nikon Metrology, Shanghai, China). The reconstructions were then analyzed with visualization software called VGStudio MAX. Straight fibers were measured by identifying their endpoints and calculating the distance between them, while curved fibers were approximated by fitting a line along the fiber axis. Their results demonstrated accurate fiber length measurements using this method [23].

Deep-learning algorithms have also been employed to extract individual fibers from lower-resolution CT scans. Konopczynski et al. utilized a deep convolutional network for segmenting fibers within the reconstructed images. This approach uses semantic segmentation, which generates a confidence map for each section of the image to identify which segments belong to each individual fiber this workflow is illustrated below in Figure 10. The analysis was performed in sections to reduce the computational resources needed to process a full-sized scan all at once. The results demonstrated improved segmentation compared to traditional skeletonization methods for these low resolution 3D scans [24].

In 2020, Henys and Čapek explored the use of multilevel machine learning to extract individual fibers from yarns in order to analyze fiber structure and mechanics [25]. Their research addressed the challenges of identifying fiber overlaps and intersections and reconnecting them within a 3D model of the fabric or yarn. They used 3D scanning to capture images of the samples, which were then converted into 3D arrays.

The images were binarized using Otsu’s method and skeletonized to produce one-pixel-width fiber segments. A neighborhood method was applied to remove any H-shaped connections between fibers. Disconnected segments were labeled and reconstructed based on their geometric relationships, their workflow is illustrated in Figure 11. These reconstructed segments were then verified using a clustering algorithm to ensure they matched the 3D scanned data. The final model achieved good accuracy, according to the authors, and was more efficient than using deep neural networks [25].

Similarly, research has explored using image analysis and modeling to characterize properties such as fiber orientation distribution, wetting, and thermal properties of non-woven fabrics. Techniques like 3D scanning and X-ray tomography are used to generate a series of 2D images for analysis. These images are binarized and segmented to either be processed under the assumption of a 2D fabric structure or used to create a 3D model of the web [26,27,28]. Since the fabrics are 3D scanned, all fiber crossovers are accurately identified to determine the final structure.

While 3D image analysis shows promise for fiber length measurement, the primary challenge is that the fibers or samples must be very small to achieve a high-resolution scan. Most research in this area focuses on fiber composite materials, where fibers, such as glass fibers, are only millimeters long. Additionally, deep-learning methods for fiber extraction require effective training with a large amount of curated and often manually labeled data. Moreover, 3D image reconstruction and analysis can demand high computational resources and be costly to perform.

3. Future Directions and Challenges

Studies thus far have relied on traditional image analysis methods with a similar workflow: binarization using standard thresholding and manual segmentation of the fibers. These methods work best for measuring straight individualized fibers such as in composites as seen in a recent study by Suárez et al. [29]. Inherently, these processes cannot identify crossed and looped fibers and therefore rely on fully individualized fiber images. In order to accurately measure a large number of fibers within a sample efficiently, there exists a gap in the market to use newer technologies to identify fiber crossovers within an image. All current advancements have centered around cotton and wool fibers using dual-beard fibrography or 3D scanning of non-woven or composite materials. Developing a new method could address these problems to create more precise and automated test method techniques.

An image segmentation method not currently used for fiber characterization involves large visual models, such as Facebook’s Segment Anything Model (SAM). For instance, Chen et al. applied the SAM for fabric defect segmentation, a challenging task due to the wide variety of defect types that need to be detected [30]. This typically requires a large amount of high-quality annotated data to accurately segment all types of fabric defects. In their study, Chen et al. significantly enhanced performance by training SAM’s general model with fabric-specific knowledge, resulting in a superior model with high accuracy and efficiency [30]. Given its success in this complex application, SAM could potentially be adapted to distinguish overlapping fibers in order to segment fibers within a 2D image for fiber length characterization.

However, instance segmentation of highly overlapping objects is a known limitation of segmentation models and object detection [31,32]. While fibers in images may not be heavily occluded, they are visually similar and therefore difficult to distinguish in a two-dimensional image. Wegmayr et al. created software to segment microfibers utilizing a visual model and deep pixel embeddings to address fiber crossovers which improves on traditional methods [33]. This further illustrates the market need for better image processing for fiber identification. In 2021, Ke et al., developed a bilayer convolutional network which uses a two-stage object detection approach to detect occluding objects as well as inferring the full position of the occluded object. They showed success with images of overlapping people, cars and animals within specified region-of-interest bounding boxes [31]. It is possible this type of modification on instance segmentation could be applied to overlapping fiber images.

Additionally, generative AI and other large image-based models have the potential to map 2D to 3D images, allowing for higher density fiber images to be modeled and characterized. Deep-learning architecture has been developed to produce three-dimensional shapes from single images. For example, in 2018, Wang et al. developed a method that progressively forms an ellipsoid shape to create detailed models [34]. More recently, in 2020, a study utilized deep learning to segment 3D tomography images of fiber composite materials [35]. It is possible this type of technology could be developed to accurately measure fiber lengths by generating a 3D model from single images of overlapping fibers.

In general, the image processing methods used to measure fiber length have limitations on identifying fiber loops or interlacements and rely heavily on manual individualization of the fibers before measurement. A key direction to move forward in this space would be to discover whether deep-learning image models could be trained to distinguish individual fibers in more complex images. Recently, fully convolutional sematic segmentation has allowed the segmentation of multiple instances of an object in images of any size [36]. Then, the process can be evaluated to see any improvements on accuracy and efficiency of the test method in comparison to manual hemp fiber measurement. A more robust methodology could allow for the commercialization and standardization of a test method that is applicable to a wide variety of fiber types.

4. Conclusions

Characterizing and measuring hemp fibers is critical for their application and processing. Although various methods exist to determine fiber length distribution and properties, most have been developed specifically for cotton or wool fibers. A summary of the current commercialized and standardized methods is provided in Appendix A. Cotton and wool fibers differ significantly from hemp and other bast fibers, which have highly variable lengths ranging from 2–10 cm to several meters—the full length of the plant. Hemp fibers are generally stronger than both cotton and wool, but they have much lower elongation and tend to be more brittle [3]. As a result, sample preparation processes designed for other fibers may not be suitable for hemp.

For example, many fiber measurement methods utilize a fiber beard, a process involving clamping and brushing the fibers, which has not been applied to hemp, likely due to its brittle nature. Similarly, methods that individualize fibers, like the AFIS, use automatic opening actions known to cause fiber breakages, potentially leading to even higher errors with hemp. Many existing methods are also designed for a narrower range of fiber lengths than what is typically observed with hemp. While manual measurement methods can be applied to any fiber, they are expensive and time-consuming, making them unsuitable for large-scale use and more appropriate for laboratory-scale projects.

Image processing presents a promising scalable and accurate measurement method for hemp fibers. The challenge lies in adapting a current method or developing a new algorithm specifically for this application. Currently, the Fibreshape (IST AG, Ebnat-Kappel, Switzerland) is the only commercial testing machine using image processing for fiber characterization, but it has a maximum fiber length measurement of 30 cm. Additionally, its algorithm does not effectively handle fiber crossovers and overlaps during analysis and is primarily used for measuring fiber fineness. While fully individualizing fibers before scanning is possible, this approach requires substantial manual effort to prevent fiber breakage.

To address these challenges, a new image analysis algorithm should be developed with a more robust segmentation process capable of distinguishing overlapping or looped fibers. This might be achieved by utilizing large image models, machine-learning techniques suited for detecting overlapping objects [37], or generating 3D models [34]. Combined with a scanner-based setup designed for optimal, even lighting, such an approach could provide a scalable solution for accurately measuring hemp fiber length.