Image recognition method of cashmere and wool based on SVM-RFE selection with three types of features

Yaolin Zhu; Kainan Liu; Meihua Gu; Kaibing Zhang; Gang Hu

doi:10.1515/aut-2025-0037

Artikel Open Access

Image recognition method of cashmere and wool based on SVM-RFE selection with three types of features

Yaolin Zhu , Kainan Liu , Meihua Gu , Kaibing Zhang und Gang Hu

Veröffentlicht/Copyright: 29. Mai 2025

Veröffentlicht von

Veröffentlichen auch Sie bei De Gruyter Brill

Manuskript einreichen Informationen für Autor*innen Erkunden Sie dieses Fachgebiet

Aus der Zeitschrift AUTEX Research Journal Band 25 Heft 1

Abstract

Cashmere and wool fibers are important raw materials in the textile industry, but their similar morphological structures make accurate distinctions challenging. Image preprocessing methods will cause some damage to the fiber contours, resulting in the loss of feature information. The keypoint features that do not require image preprocessing are added to the library of morphological and texture features. At the same time, existing methods of feature selection often ignore the relation between features and the classifier. Therefore, we propose a novel feature selection method with support vector machine-recursive feature elimination (SVM-RFE). The SVM-RFE method recursively removes the features of the least contribution to SVM classification, ultimately generating the optimal feature set. Our approach achieves a recognition accuracy of 98.06%, which is 8.34% higher than the traditional two-feature method and 6.12% higher than the three-feature method, both without feature selection. Experimental results demonstrate that keypoint features effectively compensate for the information loss caused by image preprocessing, while the SVM-RFE feature selection method can select the optimal feature subset relevant to the classifier so as to accurately distinguish cashmere and wool fibers.

Keywords: Cashmere and wool; image processing; feature selection; SVM-RFE; keypoint features; classification

1 Introduction

Cashmere and its products are recognized as high-grade textile materials because of their slenderness, lightness, and warmth [1]. However, some textile and apparel production enterprises, in pursuit of higher profits, mix wool into cashmere to reduce costs, leading to counterfeit cashmere products that disrupt the consumer market and harm public interests. Therefore, it is crucial to find a method for accurately and quickly identifying cashmere and wool fibers.

Currently, traditional methods for identifying cashmere and wool include chemical, physical, biological, image-based, and deep convolutional techniques. Chemical identification methods encompass the solution method [2] and the differential alkali solubility method [3]. Physical identification approaches include manual microscopic observation [4] and Fourier transform infrared spectroscopy [5]. Biological methods are primarily divided into two categories: protein analysis [6] and DNA analysis [7]. However, these analytical techniques often present certain limitations. For example, chemical methods tend to damage fibers and are less accurate; physical methods are heavily influenced by human factors, involve slower detection speeds, and are more subjective; and biological methods require complex procedures and incur higher costs. Although deep convolutional network methods offer potential advantages, they are generally slower, necessitate large datasets, and rely on costly experimental equipment. Recently, near-infrared (NIR) spectroscopy [8] has been widely used in cashmere and wool identification due to its fast and non-destructive characteristics. Compared to image-based methods, NIR has lower adaptability and depends on costly equipment. NIR helps to analyze the physical properties of fibers, providing valuable data that support image-based methods. In contrast, image-based methods capture the shape and texture features of fibers, resulting in lower error rates, higher accuracy, and enabling fast, precise fiber identification while overcoming the limitations of other methods.

Over the past few years, more and more scholars use digital image processing technology and machine vision technology to classify and recognize cashmere and wool fibers. By employing image acquisition equipment and processing techniques, the morphological and texture features of fibers are extracted. Subsequently, extensive testing is conducted, data are collected, and mathematical statistical models are developed through analysis and statistics to recognize cashmere and wool fibers. Feature fusion technology plays a critical role in distinguishing between these fibers. Xing et al. [9] applied a fractal algorithm to calculate box counts and information dimensions from binary images of the fibers. They also used a parallel line algorithm to measure fiber fineness. The resulting feature set was then clustered and analyzed using the K-means algorithm, achieving a recognition accuracy of 97.47%. Sun [10] proposed a classification algorithm for cashmere and wool fibers based on the LC-KSVD method. This approach combines various texture features into a multidimensional matrix and uses the LC-KSVD algorithm to distinguish between the fibers, achieving an accuracy of 91%. Zhu et al. [11] presented a classification algorithm that uses the grayscale covariance matrix and Gabor wavelet transform for texture analysis. Weights for linear summation were introduced, combining feature vectors from the spatial and frequency domains to form a six-dimensional feature vector. This vector was classified using a Fisher classifier, achieving a recognition accuracy of 93.33%. The feature fusion technique described above fuses features from different branches together and promotes feature fusion across different scales, thus improving classification accuracy. To achieve better recognition results, it is common to integrate more feature information, which provides richer data but also increases the model’s complexity and computational costs.

Feature selection addresses the impact of issues like information redundancy and overfitting on classification recognition during feature fusion. It plays a crucial role in the identification accuracy, training time, and classifier performance for cashmere and wool fibers. Zhong et al. [12] proposed using compound microscopy and projection curves for cashmere and wool identification, extracting features with recurrence quantification analysis, wavelet transform, and geometry, achieving the best accuracy of 97.47% with an SVM classifier. Xing et al. [13] proposed a fiber recognition method that extracts and analyzes morphological features. The approach makes statistical assumptions about three key characteristics: fiber height, fiber diameter, and their proportions, achieving a recognition accuracy of 94.2%. Zhu et al. [14] proposed an unsupervised feature selection method based on K-means clustering, using the DB Index, feature relevance, and cluster number to select the optimal feature subset. This subset was then input into SVM, achieving a recognition rate of 97.25%. The above feature selection algorithms effectively identify the most representative feature subset from the high-dimensional fiber features, reducing the influence of invalid and redundant features, and significantly improving classification accuracy and model stability.

Deep learning techniques are able to learn complex features from large-scale data and achieve highly automated pattern recognition and task solving. Luo et al. [15] proposed a residual network-based method for cashmere and wool fiber recognition, with an 18-layer residual network model achieving over 97.1% accuracy on a test set. Zhu et al. [16] introduced an enhanced Xception network for fibers image recognition, extracting deep features with convolutional and max pooling layers and using an improved Swish activation function. The fibers were classified with a Sigmoid classifier, achieving 98.95% accuracy. Zang et al. [17] proposed a fiber recognition method combining multi-scale geometric analysis and deep convolutional neural networks. The analysis reduces image dimensionality and redundant data, while the neural networks classify the fibers, achieving 96.67% accuracy. However, deep learning techniques perform relatively poorly in handling small sample data, and the training process of the model is relatively complex and prone to overfitting.

Many related studies on fiber features have shown that not all features contribute equally to cashmere and wool classification. While the mentioned feature selection methods effectively differentiate the fibers, the extracted feature data often contains noise and classification imbalance. Additionally, most methods only consider the relationship between fiber features and category labels without fully accounting for the interaction between the feature set and the classifier, resulting in insufficient classification accuracy. To address this, this article proposes the use of the support vector machine-recursive feature elimination (SVM-RFE) feature selection algorithm for cashmere and wool fiber classification. SVM-RFE is a high-performance wrapper method that combines SVM and a posterior term elimination search strategy. It ranks features based on the SVM weight vector and recursively selects and eliminates unimportant features, resulting in a feature set that is most relevant to the SVM classifier.

2 An overview of the cashmere and wool identification processes

In this research, we tackle the difficulty of identifying cashmere and wool fibers, a task complicated by differences in their growth environments, breeding practices, and mating conditions. To address this issue, we employ a fiber classification method based on an SVM-RFE feature selection algorithm that integrates three types of features. Morphological features and texture features are extracted through digital image processing, while keypoint descriptors are directly captured from the raw images. The SVM-RFE method is then used to select the optimal feature subset, enabling more accurate differentiation between cashmere and wool fibers.

The proposed method involves four steps, as illustrated in Figure 1. First, fiber images undergo preprocessing to eliminate backgrounds and reduce noise. Second, various features are obtained from the images: geometric morphological characteristics are calculated using chain code tracking and segmentation measurement methods, texture features are extracted from the gray-level co-occurrence matrix (GLCM), and keypoint descriptors are extracted using Shi-Tomasi corner detection and binary robust independent elementary features (BRIEF) techniques. Third, the obtained feature data are standardized using the Z-score method. Finally, feature selection is applied to the entire feature set, and the optimal subset is classified using the SVM-RFE algorithm, resulting in the highest classification performance.

Figure 1

Flow chart of research methods in this article.

3 Research methods

3.1 Dataset details

In this research, the experimental materials consisted of cashmere and wool fibers sourced from the Erdos region of Inner Mongolia in northern China. These fibers were obtained from Inner Mongolian cashmere goats and fine-wool sheep raised under controlled conditions in a semi-arid climate. To ensure sample representativeness, cashmere and wool fibers were randomly collected from individuals of varying genders and ages. The images were acquired using a scanning electron microscope (model: QUANTA-450-FEG), with the accelerating voltage set to 10 kV, a working distance of approximately 10 mm, and a magnification of 1,000×. The resulting fiber images were adjusted to a resolution of 96 dpi (both horizontally and vertically), cropped to 275 × 275 pixels, and saved to a computer. A total of 1,200 images were acquired, comprising 600 cashmere images and 600 wool images. The image acquisition equipment and raw data collection for cashmere and wool fibers are illustrated in Figure 2: (a) the image acquisition equipment diagram, (b) the acquired cashmere image, and (c) the acquired wool image.

Figure 2

Image acquisition and fiber samples: (a) image acquisition equipment, (b) cashmere, and (c) wool.

3.2 Image preprocessing of cashmere and wool fiber

The acquired cashmere and wool fiber images often contain impurities, noise, and other interfering factors that negatively impact the accuracy of feature extraction. Therefore, it is essential to preprocess these images to remove background elements and reduce noise. Initially, the grayscale values of the raw images are adjusted, and histogram equalization [18] is applied to enhance the contrast of the fiber regions and their textures. Subsequently, a global thresholding method is employed to binarize the images [19], thereby obtaining fiber skeleton images with more pronounced edges. These images are then subjected to denoising and skeleton refinement to produce clear representations of the fiber skeleton contours [20]. This refined output serves as the basis for extracting morphological features, as illustrated in Figure 3. To delineate the edges of the fiber skeleton contours, the Sobel edge detection operator is utilized. The resulting edges are then expanded to generate closed fiber regions [21]. Following this, cavities are filled, and closed regions with smaller areas are removed [22]. This process yields a fiber contour template and extracted edge contours [23]. Finally, the original image background is removed using the generated template, facilitating the extraction of texture features from the fibers, as shown in Figure 4.

Figure 3

Preprocessing steps for extracting morphological features: (a) raw image, (b) enhanced image, (c) binarized fiber image, and (d) impurity removal.

Figure 4

Preprocessing steps for extracting texture features: (a) raw image, (b) edge detection, (c) margin filling, and (d) background removal.

3.3 Standardization of cashmere and wool feature data

In this study, features are extracted from cashmere and wool fibers after image preprocessing. The extracted features include morphological features, texture features, and keypoint descriptors. An eight-dimensional set of morphological features is computed using the chain code tracking algorithm and segmentation measurement methods [24]. The GLCM approach is applied to derive 14-dimensional texture features [25]. Furthermore, the Shi-Tomasi corner detection and BRIEF methods are utilized to extract 30-dimensional keypoint descriptors from the fiber images [26,27]. In total, 52-dimensional features are obtained.

The extracted features differ in nature, size, magnitude, and availability. Directly using raw feature data can lead to the dominance of high-value features in the analysis, while the contribution of low-value features is relatively diminished. To guarantee dependable results, it is essential to standardize the extracted features, maintaining a consistent scale across the dataset. This facilitates meaningful comparisons and interpretations among different features. Therefore, the Z-score standardization method is applied to normalize the data [28]. The transformation function is provided as follows:

(1) Z = x − μ σ ,

where μ represents the sample mean, σ denotes its standard deviation, and Z corresponds to the standardized value of the sample x .

3.4 The SVM-RFE selection algorithm of three types of features

Feature selection, a key approach in dimensionality reduction, eliminates redundant or irrelevant features in data classification tasks while preserving a concise set of critical ones. This process not only decreases the computational cost of classification but also enhances the performance of machine learning methods. In this study, the feature selection algorithm based on SVM-RFE is applied. This method assesses the significance of each feature to the classifier using a weight vector derived from feature values during the training phase of the SVM model. Features with lower significance are iteratively removed through backward elimination, ultimately generating the most relevant set of features for the SVM classifier [29].

The SVM-RFE method does not directly evaluate the raw informational content of the features themselves. Instead, it assesses feature importance indirectly through their performance within the SVM model, specifically their contribution to the classification decision boundary. The advantage of this approach lies in its consideration of the features’ relevance to the classification task, rather than relying solely on their statistical properties. Consequently, SVM-RFE effectively selects the features that are most valuable for enhancing the classifier’s performance. The workflow of the SVM-RFE feature selection algorithm is shown in Figure 5.

Figure 5

Flow chart of feature selection based on SVM-RFE.

Let S = [ x m 1 ( 1 ) , .. . , x m i ( i ) , x t 1 ( i + 1 ) , .. . , x t j ( i + j ) , x p 1 ( i + j + 1 ) , .. . , x p k ( i + j + k ) ] be the wool and cashmere fiber dataset containing N initial features, where i + j + k = N , x m i denotes the ith morphological feature, x t j denotes the jth texture feature, and x m i denotes the kth keypoint description subfeature. The category label is denoted as Y = [ y ( 1 ) , y ( 2 ) , … , y ( N ) ] ，where y ( i ) ∈ [ 0,1 ] and i ∈ [ 1 , 2 , .. . , N ] . In this article, the optimal feature subset is determined by recursively eliminating the features with the least contribution in the backward elimination process. Thus, the complete process of the SVM-RFE-based feature selection algorithm is outlined as follows:

Step 1: Train the SVM using the original feature set S:

(2) L = 1 2 ∑ i = 1 N ∑ j = 1 N α i α j y ( i ) y ( j ) [ k ( x i , x j ) + λ δ i j ] − ∑ i = 1 N α i ,

(3) ∑ i = 1 N α i y ( i ) = 0 and 0 ≤ α i ≤ C , ( i = 1 , 2 , .. . , N ) .

Equation (2) needs to be minimized subject to equation (3). In both equations (2) and (3), k ( x i , x j ) represents a kernel function, δ i j is the Kronecker delta ( where δ i j = 1 if i = j , and 0 otherwise ) , and α = [ α 1 , α 2 , .. . , α N ] are the parameters to be determined. λ and C are positive constants that ensure convergence, even when the problem is not linearly separable or is poorly conditioned.

Step 2: Calculate the weight vector and feature importance using the following equations:

(4) w i = ∑ i = 1 N α i y ( i ) x ( i ) ,

(5) c i = w i 2 .

Step 3: Identify the feature with the least significant feature and remove it according to the following equations:

(6) f e = arg min ( c ) ,

(7) S = S − f e .

Step 4: Update the feature ranking list according to the following equation:

(8) R = [ f e , R ] .

Step 5: Repeat Steps 1–4 until the feature set S is empty.

It is important to note that the iteration process may also stop if the feature dataset is depleted. The stopping criterion might be based on the number of features to retain. Additionally, determining the ranking of features is essential. Deciding how many of the top-ranked features to select from the SVM-RFE ranking is a key step. To balance accuracy and robustness, we use linear SVM-RFE combined with stratified N-fold cross-validation to determine the minimal feature set that yields the highest accuracy for the SVM [30].

The following steps provide a detailed explanation of the procedure:

Step 1: Begin with the complete set of features from the cashmere and wool fiber samples from Ordos as the initial feature dataset. Use the SVM-RFE algorithm to rank the features and identify a subset, removing the least important feature. In this context, let M represent the number of initial features and M − 1 the number of features in the selected subset. The SVM employs a linear kernel.

Step 2: Divide the selected feature subset into training and testing datasets using a stratified N-fold cross-validation strategy. Calculate the classification accuracy using the same SVM model employed for recursive feature elimination.

Step 3: Update the feature subset by setting M = M − 1 and reapply the SVM-RFE algorithm to the revised subset.

Step 4: Repeat Steps 1–3 until the feature subset is empty.

Step 5: Identify the minimal feature set that yields the highest accuracy for the SVM.

4 Results and analysis

4.1 Results of fiber feature extraction

After preprocessing the fiber images, eight geometric morphological features were extracted from the binary fiber images. These features included fiber diameter, scale density, scale perimeter, scale height, scale area, diameter-to-height ratio, scale relative perimeter, and scale relative area, as presented in Table 1. To extract texture features of cashmere and wool fibers, the GLCM method was applied with experimental settings of a pixel spacing of 4 and 64 gray levels per pixel. A total of 14 feature parameters were extracted, including Mean, ASM, Entropy, Variance, Energy, Contrast, Correlation, Dissimilarity, Homogeneity, Sum Entropy, Difference Entropy, Inertia, Cluster Shade, and Cluster Prominence, in 4 directions ( θ = 0 ° , 45 ° , 90 ° , 135 ° ) . For each feature parameter, the mean across different directions was calculated, resulting in 14 texture features describing the fibers, as shown in Table 2. During the preprocessing steps for extracting morphological and texture features, separate operations were performed on cashmere and wool fibers, which may have caused minor damage to their edge contours. To mitigate this issue, the Shi-Tomasi corner detection method was employed to locate keypoints in the original fiber images. The non-maximum suppression method was then used to filter the detected keypoints, retaining the 30 most prominent ones [31]. Finally, BRIEF descriptors were utilized to extract feature descriptors from these localized keypoints, as shown in Table 3.

Table 1

Initial description of morphological characteristics

Morphological characteristics
(21.043, 0.073, 13.582, 69.251, 285.818, 1.549, 3.290, 0.645)

Table 2

Initial description of texture characteristics

Texture characteristics
(1.702, 0.649, 1.914, 15.290, 0.805, 1.513, 0.950, 0.286, 0.916, 1.914, 1.914, 5449275.632, 1037254269, 2.11 × 10¹¹)

Table 3

Initial description of keypoint descriptors

Keypoint descriptors
(12.636, −320.270, −108.782, 53.347, −31.571, −25.720, 25.046, 19.023, 17.295, −74.526, 41.836, 24.109, 50.207, 17.802, −19.548, 45.361, 20.780, −9.655, −45.723, 1.008, 19.464, 35.336, −8.006, −38.257, −3.775, −17.968, 13.685, 10.007, 6.465, −6.150)

4.2 Results of fiber optimal feature selection

After feature extraction, a total of 52-dimensional features were obtained, including morphological features, texture features, and keypoint descriptors. To determine the optimal feature subset for cashmere and wool classification, we applied the SVM-RFE algorithm with stratified fivefold cross-validation on the full dataset to rank and evaluate the features. The SVM-RFE algorithm recursively ranked the 52 features based on their contributions to a linear SVM classifier, generating a feature importance ranking, as shown in Figure 6.

Figure 6

Ranking of feature importance to the classifier.

Figure 7 illustrates the trend of average classification accuracy across different feature dimensions under fivefold cross-validation. The accuracy increases with the number of selected features, reaching a peak of 98.06% with six features (ASM, Energy, Relative Area, Density, Point_6, Mean), and subsequently decreases as more features are included. To provide a statistical basis for feature selection, we analyzed the convergence of cross-validation accuracy and defined a stopping criterion: feature selection terminates when the accuracy improvement falls below 0.3% or a decline occurs. The accuracy rises from 96.94% with four features to 97.78% with five features (an increase of 0.84%) and from 97.78 to 98.06% with six features (an increase of 0.28%), approaching the threshold and achieving the highest value. However, adding a seventh feature reduces the accuracy to 97.78%, with a continued downward trend thereafter. The trend in Figure 7 indicates that six features represent the convergence point of accuracy, beyond which additional features introduce redundancy or noise. This aligns with the theoretical foundation of SVM-RFE, which aims to retain a minimal yet highly discriminative feature subset.

Figure 7

Comparison of classification accuracy for different feature dimensions.

In this study, a linear SVM classifier was employed for the classification task based on the selected feature subset. The parameter c, which controls the penalty for misclassification, was optimized using a grid search method [32] during the stratified fivefold cross-validation process, yielding an optimal value of 11.12. With the six selected features, the linear SVM achieved an average accuracy of 98.06% on the full dataset under fivefold cross-validation, demonstrating that parameter optimization and feature selection jointly enhance model performance.

4.3 Importance of the selected features

The SVM-RFE algorithm identifies ASM, Energy, Relative Area, Density, Point_6, and Mean as the optimal feature subset, underscoring their pivotal roles in distinguishing cashmere and wool fibers. Their importance stems from their physical significance and contributions to the SVM classifier, as outlined below:

ASM (angular second moment): Derived from the GLCM, ASM quantifies the uniformity of gray-level distribution. Cashmere’s finer, smoother scales yield higher uniformity than wool’s variable texture, making ASM a key discriminator of surface differences.
Energy: Also from GLCM, energy measures local homogeneity, with higher values indicating consistent texture patterns. Cashmere’s distinct scale structure enhances energy’s sensitivity to subtle textural variations, bolstering classification accuracy.
Relative area: This morphological feature reflects the proportion of scale area to total image area. Cashmere’s compact scales contrast with wool’s larger, irregular ones, providing critical structural information that complements texture features.
Density: Defined as scales per unit length, density captures scale distribution compactness. Cashmere’s higher density, tied to its biological traits, enhances the classifier’s ability to differentiate based on scale frequency.
Point_6: Extracted via Shi-Tomasi corner detection and BRIEF, this keypoint descriptor pinpoints a discriminative local feature. Robust to preprocessing contour loss, Point_6 likely encodes a unique pattern differing consistently between cashmere and wool.
Mean: This texture feature, the average gray-level intensity, reflects optical properties influenced by scale structure and fiber diameter. It provides a global descriptor, complementing localized features.

SVM-RFE selects these features by assessing their impact on the decision boundary, quantified by the weight vector w in equation (4). Features with greater weights better separate cashmere and wool classes. Spanning texture (ASM, Energy, Mean), morphology (relative area, density), and keypoints (Point_6), this subset balances global and local properties, maximizing separability. This synergy drives the 98.06% accuracy, effectively capturing multidimensional differences between cashmere and wool fibers.

4.4 Effects of single and combined features

To explore the impact of different types of features and their combinations on the classification of cashmere and wool, we employed morphological features, texture features, and keypoint descriptors for feature selection and classification. We considered their individual use, pairwise combinations, and the combination of all three feature types. Figure 8 illustrates the classification accuracy for various feature dimensions when these features are used individually or in combination.

Figure 8

Comparison of classification performance for single and combined features.

As shown in Figure 8, there are significant differences in the recognition accuracy of cashmere and wool based on different feature types and their combinations. When using morphological features alone, the recognition accuracy is the lowest, with a maximum of only 88.6%. This is because morphological features mainly reflect the overall shape and boundary information of fibers, lacking descriptions of finer details and local structures, resulting in insufficient discriminative power. In contrast, the recognition accuracy using texture features alone is higher, reaching 93.89%, as texture features can capture subtle surface textures and structural patterns in the image, making them more effective for distinguishing materials like cashmere and wool, which have specific surface characteristics.

When combining morphological and texture features, the recognition accuracy further increases to 95%. This improvement is due to the complementary nature of the two feature types. Morphological features provide overall structural information, while texture features capture local surface details. Together, they provide more comprehensive information, enhancing the classifier’s ability to differentiate between the two materials. When morphological features, texture features, and keypoint descriptors are combined, the recognition accuracy reaches its highest value of 98.06%. Keypoint descriptors focus on local distinctive regions in the image, further enriching the description of the feature space. The combination of these three types of feature provides a comprehensive set of information, including global structure, surface details, and local significant features, enabling the classifier to better and more deeply understand the image, significantly improving recognition accuracy and classification robustness.

4.5 Evaluation of classification performance

To demonstrate the significance of feature selection algorithms in classification tasks, this study compares the performance of the feature selection method with that of a direct classification approach without feature selection. Table 4 presents the classification results on the cashmere and wool datasets, where the SVM-RFE feature selection method is compared with the traditional SVM classifier without feature selection. The results are shown for both the traditional morphological and texture features, as well as for the combined morphological, texture, and keypoint descriptor features proposed in this study.

Table 4

Performance of the feature selection algorithms

Feature selection	Types of feature	Accuracy (%)	Precision (%)	Recall (%)	F1-score (%)
SVM	Morphological + texture	89.72	88.89	90.91	89.89
SVM	Morphological + texture + keypoint	91.94	91.67	92.18	91.92
Ours (SVM-RFE)	Morphological + texture + keypoint	98.06	97.27	98.89	98.07

According to the results in Table 4, whether using traditional morphological and texture features or the proposed combination of morphological, texture, and keypoint descriptors, the classification performance of the SVM-RFE feature selection method is significantly superior to that of the traditional SVM method without feature selection. Specifically, when using morphological and texture features, the accuracy of the traditional SVM method is 89.72%, which increases to 91.94% when keypoint descriptors are incorporated. In contrast, the SVM-RFE method achieves an accuracy of 98.06% when all three feature types are combined, far exceeding the highest accuracy of the traditional SVM method. These results demonstrate that the SVM-RFE feature selection method effectively optimizes the feature space by eliminating redundant and irrelevant features, thereby greatly enhancing the classifier’s overall performance.

Our method demonstrates sensitivity to the type of classifier used. As shown in Table 5, the SVM-RFE method employed in this study achieves the highest classification accuracy, reaching 98.06%. In contrast, KNN-RFE shows the lowest performance, likely because the KNN algorithm is less sensitive to high-dimensional data. The feature selection process may remove information that is useful to KNN, resulting in decreased classification performance. Decision trees and random forests, as tree-based classifiers, generally perform well with high-dimensional data and feature selection. However, due to their tree structure, they may not fully exploit the feature information within the data. Therefore, the SVM-RFE approach adopted in this study is simple, easy to implement, and effectively addresses the cashmere and wool classification problem. The accuracy of our conclusions has been verified through comparative experiments with other classifiers.

Table 5

Performance of different classifiers combined with RFE

Classifier	Accuracy (%)	Precision (%)	Recall (%)	F1-score (%)
KNN-RFE	76.39	75.96	77.22	76.59
Decision Tree-RFE	94.17	93.92	94.44	94.18
Random Forest-RFE	96.11	96.63	95.56	96.09
Ours (SVM-RFE)	98.06	97.27	98.89	98.07

Table 6 presents an evaluation of the proposed methods using various fiber mixing ratios. The highest identification accuracy is achieved when the sample sizes of cashmere and wool are equal at a 1:1 ratio. This balanced sample size facilitates better characterization of the distinct fibers. The proposed method demonstrates stable performance, with accuracy ranging from 96 to 98% across different cashmere and wool blending ratios.

Table 6

Accuracy of identification with different sample sizes

The number of samples cashmere (wool)	Ration cashmere (wool)	Accuracy (%)
100/600	1:6	96.38
200/600	2:6	96.67
300/600	3:6	96.94
400/600	4:6	97.22
500/600	5:6	97.78
600/600	6:6	98.06
600/500	6:5	97.78
600/400	6:4	97.50
600/300	6:3	97.22
600/200	6:2	96.94
600/100	6:1	96.67

4.6 Comparison experiments with current fiber identification methods

To ensure comparability, this study compares its proposed method with several existing techniques for analyzing cashmere and wool fibers, as shown in Figure 9. The method adopted here uses SVM-RFE for feature selection, which has demonstrated significantly high recognition accuracy. This approach evaluates and integrates features related to the morphology, texture, and keypoints of cashmere and wool fibers. By focusing on these crucial attributes, the method generates a feature set highly relevant to the classification task, resulting in more precise and efficient classification and identification. The experimental results further highlight the effectiveness of this method, showing a substantial improvement in recognition accuracy compared to existing techniques.

Figure 9

Comparison of various recognition methods.

5 Conclusion

The article proposes a feature selection method based on SVM-RFE, incorporating three types of features for the classification and identification of cashmere and wool fibers. Initially, the images of cashmere and wool fibers are preprocessed using two distinct methods to generate the fiber scale skeleton map and the target fiber region map. Subsequently, morphological features and texture features are extracted, along with keypoint descriptors derived from the original fiber images, resulting in a total of 52-dimensional features. The SVM-RFE feature selection technique is then employed to rank these features based on their relevance to the SVM classifier. An optimal subset of features is chosen from this ranked list, and classification predictions are made. The results of the experiments demonstrate that this method achieves a high recognition rate, with a noticeable improvement over other algorithms.

Despite these advantages, this approach also faces limitations. The sample size of 1,200 images may not capture the full fiber variability and may reduce generalizability. Relying on high-quality images may lead to degraded performance when dealing with noisy or low-resolution data. Additionally, it focuses on cashmere and wool, which limits its applicability to other fiber types. These limitations highlight opportunities for improvement.

To address these challenges, future research will explore advanced feature selection and unsupervised methods to overcome label scarcity and costs. Expanding the dataset with diverse modified fibers will enhance generalizability and robustness, particularly under varying image conditions. These efforts aim to strengthen the method’s practical utility for textile quality control and beyond.

Acknowledgments

The authors express their gratitude to the Erdos Textile Research Center for providing the cashmere and wool fiber samples and to the Xi'an Polytechnic University Laboratory for granting access to the scanning electron microscope used in this study.

Funding information: The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the National Natural Science Foundation of China (6257072408), the Key Research Project of Yulin Science and Technology Bureau (2024-CXY-159), the Key Program for Basic Research of Natural Science by the Shaanxi Provincial Science and Technology Department (No. 2023-JC-ZD-33), the Key Technology Research Project of Industrial Chain by the Xi’an Science and Technology Bureau (No. 23ZDCYJSGG008-2023), and the “Chief Scientist-Chief Engineer” Team Construction Project for the Intelligent Identification of Fiber Components in Woolen Cold-proof Wadding (No. YLKG-2023-04).
Author contributions: Kainan Liu: Methodology, Writing, Editing. Yaolin Zhu: Conceptualization, Review. Meihua Gu: Software. Kaibing Zhang:Investigation. Gang Hu: Sample collection.
Conflict of interest: The authors declared no potential conflicts of interest regarding the research, authorship, and/or publication of this article.
Ethical approval: The conducted research is not related to either human or animals use.
Data availability statement: The datasets generated and/or analyzed during the current study are available from the corresponding author on reasonable request.

References

[1] Karthik, T., Rathinamoorthy, R., Ganesan, P. (2015). Sustainable luxury natural fibers – production, properties, and prospects. Handbook of Sustainable Luxury Textiles and Fashion, 1, 59–98.10.1007/978-981-287-633-1_4Suche in Google Scholar

[2] Atav, R., Türkmen, F. (2015). Investigation of the dyeing characteristics of alpaca fibers (Huacaya and Suri) in comparison with wool. Textile Research Journal, 85(13), 1331–1339.10.1177/0040517514563727Suche in Google Scholar

[3] Tang, J., Zhao, S., Yin, H. (2014). Identification methods descriptions of wool and cashmere. Wool Textile Journal, 42(7), 48.Suche in Google Scholar

[4] Tonetti, C., Varesano, A., Vineis, C., Mazzuchetti, G. (2015). Differential scanning calorimetry for the identification of animal hair fibres. Journal of Thermal Analysis and Calorimetry, 119, 1445–1451.10.1007/s10973-014-4247-8Suche in Google Scholar

[5] Xu, W., Xia, J., Min, S., Xiong, Y. (2022). Fourier transform infrared spectroscopy and chemometrics for the discrimination of animal fur types. Spectrochimica Acta Part A Molecular & Biomolecular Spectroscopy, 274, 121034.10.1016/j.saa.2022.121034Suche in Google Scholar PubMed

[6] Li, S., Zhang, Y., Wang, J., Yang, Y., Miao, C., Guo, Y., et al. (2016). Combining untargeted and targeted proteomic strategies for discrimination and quantification of cashmere fibers. PLoS One, 11(1), e0147044.10.1371/journal.pone.0147044Suche in Google Scholar PubMed PubMed Central

[7] Zhang, X., Wu, X., Yang, H., Zheng, H., Zhou, Y. (2023). Identification of cashmere and wool by DNA barcode. Journal of Natural Fibers, 20(1), 2175100.10.1080/15440478.2023.2175100Suche in Google Scholar

[8] Chen, X., Lan, Q., Zhu, Y. (2024). Non-destructive identification of wool and cashmere fibers based on improved LDA using NIR spectroscopy. AUTEX Research Journal, 24(1), 20230017.10.1515/aut-2023-0017Suche in Google Scholar

[9] Xing, W., Xin, B., Deng, N., Chen, Y., Zhang, Z. (2019). A novel digital analysis method for measuring and identifying of wool and cashmere fibers. Measurement, 132, 11–21.10.1016/j.measurement.2018.09.032Suche in Google Scholar

[10] Sun, C. (2021). Image classification of cashmere and wool fiber based on LC-KSVD. Journal of Physics: Conference Series, 1948(1), 6.10.1088/1742-6596/1948/1/012010Suche in Google Scholar

[11] Zhu, Y., Huang, J., Wu, T., Ren, X. (2021). Identification method of cashmere and wool based on texture features of GLCM and gabor. Journal of Engineered Fibers and Fabrics, 16, 1558925021989179.10.1177/1558925021989179Suche in Google Scholar

[12] Zhong, Y., Lu, K., Tian, J., Zhu, H. (2017). Wool/cashmere identification based on projection curves. Textile Research Journal, 87(14), 1730–1741.10.1177/0040517516658516Suche in Google Scholar

[13] Xing, W., Liu, Y., Deng, N., Xin, B., Wang, W., Chen, Y. (2020). Automatic identification of cashmere and wool fibers based on the morphological features analysis. Micron, 128, 102768.10.1016/j.micron.2019.102768Suche in Google Scholar PubMed

[14] Zhu, Y., Wang, X., Gu, M., Hu, G., Li, W. (2024). Application of unsupervised feature selection in cashmere and wool fiber recognition. Journal of Natural Fibers, 21(1), 2311306.10.1080/15440478.2024.2311306Suche in Google Scholar

[15] Luo, J., Lu, K., Chen, Y., Zhang, B. (2021). Automatic identification of cashmere and wool fibers based on microscopic visual features and residual network model. Micron, 143, 103023.10.1016/j.micron.2021.103023Suche in Google Scholar PubMed

[16] Zhu, Y., Huang, J., Li, Y., Li, W. (2022). Image identification of cashmere and wool fibers based on the improved Xception network. Journal of King Saud University-Computer & Information Sciences, 34, 9301–9310.10.1016/j.jksuci.2022.09.009Suche in Google Scholar

[17] Zang, L., Xin, B., Deng, N. (2022). Identification of wool and cashmere fibers based on multiscale geometric analysis. The Journal of The Textile Institute, 113(6), 1001–1008.10.1080/00405000.2021.1914399Suche in Google Scholar

[18] Stark, J. A. (2000). Adaptive image contrast enhancement using generalizations of histogram equalization. IEEE Transactions on Image Processing: A Publication of the IEEE Signal Processing Society, 9(5), 889–896.10.1109/83.841534Suche in Google Scholar PubMed

[19] Koide, H., Kim, C., Shono, K. (1991). Facsimile image processor using threshold-adaptive binarization technique. ICVC: International Conference on VLSI and CAD, Vol. 2.Suche in Google Scholar

[20] Kumar, S. (2013). Image denoising based on gaussian/bilateral filter and its method noise thresholding. Signal, Image and Video Processing, 7(6), 1159–1172.10.1007/s11760-012-0372-7Suche in Google Scholar

[21] Morales, O., García, P., Reyes, F., Lorenzo, G., Juan, V. (2012). Experiences in the implementation of the morphological erosion and dilation operations for binary images using adaptive neighborhoods. Ingeniería Electrónica, Automática y Comunicaciones, 33(2), 34–41.Suche in Google Scholar

[22] Wang, T., Chen, G., Guan, H. (2010). An improved filling algorithm for image with complicated closed edges. Applied Mechanics and Materials, 893(20–23), 894–897.10.4028/www.scientific.net/AMM.20-23.894Suche in Google Scholar

[23] Han, Y., Jun, Y., Fabian, R. (2006). Segmentation-assisted edge extraction algorithms for SEM images. Proceedings of SPIE, Bellingham, Washington, USA, pp. 458–467.Suche in Google Scholar

[24] Wang, J., Song, W., Zhao, L., Wang, W. (2014). Application of improved freeman chain code in edge tracking and straight line extraction. Journal of Signal Processing, 30(4), 422–430.Suche in Google Scholar

[25] Garg, M., Dhiman, G. (2021). A novel content-based image retrieval approach for classification using GLCM features and texture fused LBP variants. Neural Computing and Applications, 33(4), 1311–1328.10.1007/s00521-020-05017-zSuche in Google Scholar

[26] Bansal, M., Kumar, M., Kumar, M., Kumar, K. (2020). An efficient technique for object recognition using Shi-Tomasi corner detection algorithm. Soft Computing, 25(6), 4423–4432.10.1007/s00500-020-05453-ySuche in Google Scholar

[27] Wissal, M., Basel, S., Atef, H., Zouhour, D., Khaled, B. (2020). Comparative study of feature detector and descriptor methods for registration. Twelfth International Conference on Machine Vision, Vol. 11433.Suche in Google Scholar

[28] Fei, N., Gao, Y., Lu, Z., Xiang, T. (2021). Z-score normalization, hubness, and few-shot learning. Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 142–151.10.1109/ICCV48922.2021.00021Suche in Google Scholar

[29] Piyushkumar, A., Jagath, C. (2010). SVM-RFE with MRMR filter for gene selection. IEEE Transactions on Nanobioscience, 9(1), 31–37.10.1109/TNB.2009.2035284Suche in Google Scholar PubMed

[30] Wang, C., Xiao, Z., Wu, J. (2019). Functional connectivity-based classification of autism and control using SVM-RFECV on rs-fMRI data. Physica Medica, 65, 99–105.10.1016/j.ejmp.2019.08.010Suche in Google Scholar PubMed

[31] Hong, L., Ting, W., Qian, Z., Hui, Z. (2022). Non-maximum suppression performs later in multi-object tracking. Applied Sciences, 12(7), 3334–3334.10.3390/app12073334Suche in Google Scholar

[32] Li, S., Chang, J., Chu, M., Li, J., Yang, A. (2022). A blast furnace coke ratio prediction model based on fuzzy cluster and grid search optimized support vector regression. Applied Intelligence, 52, 13533–13542.10.1007/s10489-022-03234-8Suche in Google Scholar

Received: 2024-12-22

Revised: 2025-04-10

Accepted: 2025-04-15

Published Online: 2025-05-29

This work is licensed under the Creative Commons Attribution 4.0 International License.

Artikel in diesem Heft

https://doi.org/10.1515/aut-2025-0037

Schlagwörter für diesen Artikel

Cashmere and wool; image processing; feature selection; SVM-RFE; keypoint features; classification

Creative Commons

BY 4.0