Classification of cotton content in fabrics

Classifying fabrics after their cotton content is a task useful in recycling as well as quality control [4]. But doing this task manually is time-consuming and cost-intensive [4]. A common approach to solve these problems for the task of classifying cotton content is NIR spectrography [6]. While this method is deemed effective, it is cost-intensive and creates complex datasets which are complicated to analyze [6]. A lot of related works try to analyze NIR spectrography data with regression models or deep learning (DL) methods. Some works in this field also focus on different tasks, such as classifying cotton-polyester blends [8]. Other research also does not focus on classifying cotton content in fabrics, but detecting faults in cotton lint, a waste product not usable for recycling [12]. Doing the task of classifying cotton content from fabrics visually is a way to automate the analog process and make it more time- and cost-effective [4]. Despite this, visual approaches to cotton content classification are rare.

Deep learning in classifying cotton-content

The amount of literature on the subject of classifying cotton content using deep learning is not extensive. To our knowledge, we are the first to work with the CottonFabricImageBD dataset [4], which could be rooted in its novelty. But there is a larger amount of literature on classifying fabric contents with DL methods using other datasets. We also make a focus of the literature on near- and mid-infrared spectrography images. The literature on purely visual methods is limited.

Visual approaches.

Other research focuses on visual approaches rather than NIR spectrography; relevant examples are outlined in the following. With the use of a multi-class SVM, Li et al. [12] wanted to classify foreign fibers in cotton lint. Trying to classify six different classes, namely hair, black plastic film, red cloth, hemp rope, red polypropylene, and black feather. Comparing different SVM approaches separately, such as one-against-all decision-tree based MSVM, one-against-one voting-based MSVM, and one-against-one directed acyclic graph MSVM, they reach an accuracy of 93.57% with a one-against-one voting-based approach and 92.34% with a directed acyclic graph approach. Only these two approaches reach their accuracy requirements, with the one-against-one directed acyclic graph MSVM being the fastest in classification speed of all three approaches. Closer related to our work is the study of Islam et al. [14] who did a cotton-content prediction from fabric images, using a CNN with transfer learning and reached an RSME of 7.56%. They did not use a cross-validation approach in combination with a photometric stereo sensor dataset. We address these limitations by using a stratified 5-fold cross-validation, as well as a peer-reviewed dataset containing images captured with a normal camera. We also do not focus on prediction, but on the classification of the cotton contents. Table 1 provides a short overview of the described related work.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Overview of related work for fabric classification.

https://doi.org/10.1371/journal.pone.0346583.t001

Model architecture

Hybrid architecture.

We used transfer learning in comparison to training an entire model from scratch. As mentioned, the base models used in this study are DenseNet121 and a Swin Transformer V2. DenseNet121 was chosen over other backbones because it is particularly well suited for representing dense features and thus, for example, subtle tissue structures across all layers. This is not necessarily the case to the same extent with other backbones, such as those that work heavily with residual connections. The Swin Transformer V2 was selected because its shifting window attention makes it particularly well suited for processing spatially repetitive features, which is less the case with transformers that use token mixing [23]. We modified DenseNet121 with DConv layers in between DenseBlocks three and four (see Fig 3), as well as an AFPN. Each of the two architectures independently extracts relevant features from the input images. The outputs of both architectures are then combined. The initial two-dimensional results of both models are simplified to a one-dimensional feature map by adaptive averaging and flattening. These 1D vectors are then processed by batch normalization. Mathematically, batch normalizations work applying the Eq. (1) and (2) [16]. Eq. (1) displays the normalization step, while Eq. (2) displays the following step of scaling and shifting. Here x_i are activations, B is the batch size, and is the batch variance, while is a constant added for numerical stability. and are to be learned parameters, and BN is the batch-normalization transform. Finally, y_i are the scaled and shifted values [16]:

(1)(2)

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Visualization of the DenseNet121 architecture, in which an AFPN and a DConv layer is integrated between DenseNet blocks three and four.

https://doi.org/10.1371/journal.pone.0346583.g003

These one-dimensional outputs of both are then fused via the concatenation function and batch normalization is applied again. This second normalization re-aligns the feature statistics after fusion, preventing scale discrepancies between the CNN and Transformer representations. After the fusion, a new vector is created where and are the two output vectors of DenseNet121 and the Swin Transformer V2, n is the length of the respective vector. The Swin Transformer V2 combines different areas of the input image to recognize how these are connected. Thereby, local information is combined with global information [15]. This helps in our case to detect different fibers and connect this information to the overall structure of the fabric, enabling the network to better detect the cotton content of the respective fabric. Through the use of two FC layers after the concat fusion, we aimed to connect the extracted differences in order to enhance classification accuracy. This is needed especially for classifying cotton contents with small percentage differences and by that enhance the model’s classification capacity, a task for which using more FC layers is considered useful [17]. Two fully connected layers were selected empirically, as preliminary tests had shown that a single FC layer resulted in lower validation performance. An additional FC layer was not used due to the increasing number of parameters and complexity. The number of neurons in the first FC is determined by hyperparameter tuning, the number of neurons for the second FC is 13, the number of classes.

Deformable convolution layer.

DConv layers can adapt to specific structures or objects in the data, making them useful for our task of detecting cotton content, as the differences here are based on the structures of the fabrics. They add learnable offsets to the sampling positions within the filter, making them no longer fixed (see Fig 4) [19]. With these offsets, which are calculated through additional convolution layers, the network can adapt and detect specific objects or structures in the data [20]. This is especially interesting when detecting cotton content, as different structures of fibers are a differentiating factor. They work by applying Eq. (3), where p₀ is the processing pixel, is the deformable convolutions offset, and R is sampling the regular grid of convolution:

(3)

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Visualization of deformable convolution with the example of a 3×3 convolution filter.

(1) normal convolution, (2) deformable convolution.

https://doi.org/10.1371/journal.pone.0346583.g004

Adaptive feature pyramid network.

Feature Pyramid Networks (FPNs) combine features from different levels of the resolution through addition, making the network better at considering features from global as well as detailed levels [21]. AFPNs are, like FPNs used to represent extracted features at multiple levels of the resolution, where higher feature maps are merged with with lower levels, being different in assigning adaptive weights to each of these feature maps before merging, to decide which features are included how much in the final feature map [22]. AFPNs are visualized in Fig 5. They help make the network consider features extracted at different levels and therefore focus more on overall structures.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Illustration of an AFPN.

https://doi.org/10.1371/journal.pone.0346583.g005

Swin transformer V2.

As it is one part of our hybrid architecture, the Swin Transformer V2 is an essential part of our approach. The Swin Transformer uses hierarchical feature maps by merging image patches [23]. With this technique, global features are combined with local features, representing information from both levels [15]. As Swin stands for Shifted WINdows, the shift is a key design element of the Swin Transformer [23]. Here, the window partition between two consecutive layers shifts, making the windows bridge edges of the ones from the preceding layer, which enhances the modeling power of the Swin Transformer significantly, as well as making it more efficient regarding latency [23]. The Swin Transformer down-samples the resolution by 2x and thereby reduces the amount of tokens equally in each layer as the network gets deeper to represent features from each layer hierarchically [23]. Additionally, the shifted window approach of the Swin Transformer has a much lower latency than other sliding window approaches and linear computational complexity to image size, making it suitable for a variety of tasks [23]. How the Swin Transformer functions is visualized in Fig 6. In this figure, displayed Swin Transformer block replaces the multi-head self-attention (MSA) in a transformer block with a module based on shifted windows, which is followed by a 2-layer Multi-Layer Perceptron (MLP) with Gaussian Error Linear Unit (GELU) non-linearity in between and a LayerNorm layer applied before, as well as a residual connection applied after each MSA and MLP module [23]. The Swin Transformer uses a patch size of 4 × 4, because of which the feature dimension of each patch is 4 × 4 × 3 = 48, and on this raw-valued feature, a linear embedding layer, denoted as C, is applied to project it to an arbitrary dimension [23]. The second version (V2) of the Swin transformer enhanced training stability and overall enhanced accuracy, due to a new normalization configuration which produces milder activation values as well as scaling up the model capacity [15]. The Swin Transformer is generally well-suited for classification tasks [23]. In our use case, it is additionally helpful to enable the network to detect local features such as different fabric structures, as well as the number of those fibers, a global feature.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Visualization of the Swin Transformer architecture.

https://doi.org/10.1371/journal.pone.0346583.g006

Process of training

The entire training process is shown in Fig 7. Before model training, a stratified 5-fold cross-validation was applied using the scikit-learn library [24] to divide the dataset into five equally sized subsets. This method was chosen to ensure that each fold maintains the original class distribution, unlike standard k-fold cross-validation, which may produce biased results due to class imbalance. In each iteration, four folds (80%) are used for training, and one fold (20%) is held out for testing. This process is repeated five times, rotating the test fold each time to ensure robust performance evaluation. Within the training set of each fold, 10% is further reserved for validation and hyperparameter tuning. All images were resized from 900×1200–256×256 pixels using bilinear interpolation with anti-aliasing to avoid aliasing artifacts, and subsequently normalized using the ImageNet mean and standard deviation for the three RGB channels. This input size ensures artifact-free patch partitioning in Swin Transformer V2 and provides slightly higher spatial detail compared to 224×224, allowing fine fabric textures to be better preserved. To reduce overfitting and improve generalization, the training data is augmented using RandomRotation(±10°) and RandomResizedCrop(scale=(0.9, 1.0)) as well as RandomAffine(translate=(0.1, 0.1)). These transformations account for variability in image scale and position, making the model more robust to input variations.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. Training and Evaluation approach: The dataset is split into a training set and a test set.

Data augmentation is applied to the training set, followed by model training using transfer learning and fine-tuning. Finally, the resulting model is evaluated on the test set.

https://doi.org/10.1371/journal.pone.0346583.g007

The techniques used were chosen conservatively in order to preserve the texture-relevant characteristics of the fabric, which are crucial for classifying the cotton content. Slight geometric transformations are used. In order not to distort the structure of the fabric too much and not to reduce the model performance, the parameter interval was limited to smaller values, these smaller intervals are also based on previous work. All transformations are applied on-the-fly during training. The chosen augmentations aimed to maintain texture features important for identifying cotton content. These augmentations do not change the weave structure, but rather simulate realistic variations in recording conditions. This preserves the fiber orientation, weave pattern and local texture. These data augmentation techniques are only applied to the training data. The validation and test data are not augmented. Each model is trained for a maximum of 100 epochs with a batch size of 16. The training consists of two phases: transfer learning and fine-tuning. To ensure stable optimization and comparability, the batch size was fixed to 16 during transfer learning, since only the classification head was trained. During fine-tuning, we explored batch sizes of 8, 16, and 32 to assess their influence on convergence and generalization when all layers were unfrozen. We selected a batch size of 16, as smaller batches can lead to unstable gradients, while a batch size of 32 or more increases memory consumption. During transfer learning, the pretrained feature extraction layers (based on ImageNet weights [25]) are frozen, while the fully connected layers remain trainable. In the fine-tuning phase, the entire model is updated. For both training phases, we used Bayesian optimization based on the Tree-Structured Parzen Estimator (TPE, Optuna) for sequential hyperparameter selection. The following hyperparameters are used: dropout, learning rate, weight decay fully connected layer units, batch size, and the optimizer. The corresponding interval can be seen in Table 2. For each fold, 20 trials are conducted and the configuration that yields the lowest validation loss is selected. At each optimization step, optuna saves the lowest validation loss. At the end of the hyperparameter optimization, all 20 validation losses are compared, and the lowest one is output with the corresponding parameter combination. Each trial is allowed to run up to 100 epochs. To avoid overfitting and reduce computational overhead, early stopping is implemented, terminating training if the validation loss does not decrease over ten consecutive epochs. Early stopping does not end the entire optimization process, but only the training process of the respective trial. During training, a constant learning rate was used and no learning rate schedule. After identifying the optimal hyperparameters, a final transfer learning model is trained using the same configuration as during optimization. The model with the lowest validation loss is saved and used as the base for the subsequent fine-tuning phase. During fine-tuning, the hyperparameters (see Table 2) are optimized again. To accomplish this, the entire model is unfrozen to allow for additional domain adaptation. Again, the best-performing model based on the lowest validation loss is stored. CrossEntropy is used exclusively as the loss function for the entire training process. Although additional distance metrics such as RMSE and Mean Absolute Error (MAE) are used for evaluation in this study, no other loss function is used. These metrics are calculated exclusively after training based on the predicted classes. After training, the final model is evaluated on the 20% test set using the following performance indicators such as accuracy, balanced accuracy, RMSE, MAE, true positive rate (sensitivity or recall), true negative rate (Specificity), positive predictive value (precision), negative predictive value, Cohen’s Kappa, and F1-score. The model outputs logits, and class predictions are obtained via argmax. The results are reported based on the model’s native outputs, without a post-hoc calibration. All these metrics are computed for each fold, stored, and finally averaged across all five folds to provide a comprehensive overview of model performance.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Overview of the hyperparameters used for hyperparameter tuning. TL = Transfer learning; FT = Fine-Tuning.

https://doi.org/10.1371/journal.pone.0346583.t002

Evaluation metrics

For the evaluation of our model, we applied the metrics (Balanced) Accuracy, true positive rate (sensitivity or recall), Positive predictive value (precision), true negative rate (Specificity), negative predictive value, and Cohen’s Kappa, as well as the F1-Score. These metrics relate to the multiclass case. Weighted metrics provide a more realistic assessment of performance when the dataset is unbalanced. When the dataset is balanced, the result corresponds to that of the macro metric. In addition to these classification metrics, we also report RMSE and MAE. While many related studies treat cotton-content prediction as a regression task using distance-based metrics such as RMSE, the present dataset defines fixed percentage levels as discrete classes, reflecting industrial practice where cotton textiles are produced with standardized compositions. A classification formulation is therefore appropriate, yet the numerical distances between classes remain relevant. Consequently, RMSE and MAE serve as complementary metrics that quantify how far the predicted percentage class deviates from the true value on this ordered scale, with RMSE emphasizing larger deviations and MAE capturing the average absolute error. The overall effectiveness of a model is displayed through the accuracy [26]. The balanced accuracy is used to avoid inflated accuracy scores with imbalanced datasets [24]. The accuracy is calculated by Eq. (4), the balanced accuracy by Eq. (5). Balanced accuracy is calculated as the arithmetic mean of the true positive rate TPR_i for each class i∈{1,..,N} [24]:

(4)

(5)

RMSE quantifies how far predicted values deviate from the true values on average. It is defined as the square root of the mean squared error, where the squared differences between predictions and ground-truth values across all samples are averaged [27]. The MAE captures how much predictions differ from the true values on average, based on absolute deviations [27]. In our case, we use the percentage values of the cotton content associated with each class and compute the deviation between predicted and true percentages. This provides a reliable measure of the magnitude of misclassification on the underlying ordinal scale, as it explicitly accounts for the numerical distances between classes. The RMSE is given by:

(6)

(7)

The true positive rate (TPR) shows how accurately the model can correctly classify the positive class. When it is maximized, the likelihood of correctly identifying true members of the positive class is increased [28]. For multiclass tasks, the TPR is computed as a weighted average, with each class weighted according to its number of samples [29]. The equation is shown in 8. The positive predictive value (PPV) indicates the proportion of positive predictions that correctly match true positive instances and, by that, displays the prediction accuracy for the positive class [28]. Both are calculated using the Eq. (8) and (9), where n_i represents the number of samples per class and S denotes the total number of samples:

(8)

(9)

The true negative rate (TNR) is analogous to the TPR, the measure of how effectively the negative classes are classified as such [26]. In the multiclass case, TNR is computed per class using a one-vs-rest strategy and aggregated using negative-support weights TN_i + FP_i per class. Similar to its positive equivalent, the negative predictive value (NPV) is the ratio of correctly classified negative samples to all samples classified as negative [30], in the multiclass case, NPV is computed per class using a one-vs-rest strategy and aggregated with TN_i + FN_i per class. Analogue to their positive counterparts, these are displayed below in the Eq. (10) and (11):

(10)

(11)

Cohen’s kappa describes the reliability of a model by measuring the agreement between two judgments. It ranges from −1–1, where a Cohen’s kappa of −1 indicates complete disagreement, and a Cohen’s kappa of 1 signifies perfect agreement. Cohen’s kappa is calculated using the Eq. (12), where p₀ is the number of times the predicted value matches the actual value and p_c is the proportion of matches that would be expected by coincidence [31]:

(12)

The F1-score is computed as the harmonic mean of precision and recall, ranging from 0 to 1, where 0 indicates that all positive samples are misclassified and 1 represents perfect classification [32]. This equation for multiclass [33] is visualized below as Eq. (13):

(13)

Dataset

To evaluate our novel hybrid approach, we use the CottonFabricImageBD dataset [4]. In total, 1,300 original fabric images are provided, with a typical image resolution of about 900 × 1200 pixels. The images are labeled in thirteen classes representing the following cotton contents: 30%, 40%, 50%, 53%, 58%, 60%, 63%, 65%, 66%, 80%, 95%, 98%, and 99%. Each class contains 100 images. Fig 8 shows examples of images at different cotton content levels to compare the images visually. The images in each class vary in colors and patterns. The authors sourced the fabrics from different sources in Bangladesh, including fabric stores and clothing manufacturers. Images of the fabrics were captured with a 48 MP (f/2.0, (wide), 1/2.0″, 0.8 m, PDAF) camera. The images were taken under consistently good lighting conditions using automatic exposure. They therefore represent a semi-controlled environment, as would be expected in real production results. Under the supervision of two textile engineering experts, they used a thread-counting machine to assess the cotton percentage of each fabric. In this process, the authors utilized the counting grid of the thread-counting machine and calculated the cotton content using the traditional thread count formula, a common practice in the textile industry. The measurement accuracy or error margin of the thread-counting machine is not mentioned. Through this process, 100 images in each class were sourced, resulting in a total of 1,300 images.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. Example images from the dataset for each class.

Shown are two images per cotton content class (13 classes in total) [4].

https://doi.org/10.1371/journal.pone.0346583.g008

Setup

For training and testing the architecture, a NVIDIA L40S GPU with 48 GB of memory, with PyTorch 2.5.0 is used. Furthermore, Python version 3.11.7 and CUDA version 12.4.1 were used. The architecture was trained with the best optimizer found in the hyperparameter tuning for a maximum of 100 epochs in various phases, such as transfer learning and fine-tuning. To identify the optimal parameters for transfer learning and fine-tuning, hyperparameter tuning was performed in 20 trials using Optuna (version 4.2.1). To avoid overfitting and save computation time, the callback function early stopping with a delta = 0.001 was used, which stops the training after 10 consecutive epochs in which the validation loss has not decreased. Scikit-learn (version 1.5.2) was utilized for stratified cross-validation and the computation of performance indicators. Throughout the entire training and validation process, the images were converted to a resolution of 256×256 pixels.

Contextualization of the results

To contextualize the achieved RMSE of 14.01%, the proposed hybrid architecture is compared with the baseline models. DenseNet121 yields an RMSE of 17.87% (+3.86 pp), while the Swin Transformer also exhibits a higher RMSE (+2.34 pp), noting that lower RMSE values indicate better performance. The values in parentheses indicate the deviation in percentage points from the hybrid architecture.

In [7], the fiber content of blended fabrics is predicted using NIR and CNNs. The study aims to improve the accuracy and robustness of composition analysis. Several methods are evaluated, showing a considerable variation in RMSE values for cotton. Random Forest achieves 17.70%, an encoder-based model 18.20%, and the ResNet model reaches an RMSE of 13.70%, the best model reaches an RMSE of 11.70%. In [10], Paz and Sousa investigate the ability of NIR and MIR to determine the cotton content in blended fabrics. Using the NIR and MIR method, an RMSE of 7.8% and 8% was achieved on textiles with cotton, polyester and other material. For textiles with cotton and polyester RMSE of 3.6% (NIR) and 6.5% (MIR) are achieved. Further studies that also used NIR, such as Xia et al. [8] and [11], achieved RMSE values of 0.65% and 2.10%, respectively. However, a direct comparison between the studies that used NIR and our study is hardly possible, as the underlying methodology differs significantly. While spectroscopy captures information that is directly linked to the cotton content via wavelength absorption, RGB imaging is based solely on the visual appearance of the fabrics. These measurement methods therefore encode less information than spectroscopy-based methods and are thus intrinsically limited. Even if a perfectly functioning model were available, it would not be able to distinguish between fabrics with different cotton content but identical visual appearance. This means that there is an inherent accuracy ceiling for these visual methods, and the RMSE of NIR approaches should therefore be understood as a measurement-driven accuracy ceiling rather than competition for visual detection methods.

To our knowledge, the work by Islam et al. [14] represents the only study that adopts a visual approach to classify cotton content in fabrics and therefore provides a comparison to our method. With the VGG16 model and transfer learning, they achieve an RMSE of 7.56%. However, this study addresses a different experimental setting and the RMSE value is therefore not directly comparable to our setting. Although our RMSE of 14.01% did not surpass their RMSE of 7.56%, this does not indicate lower performance but results from a more constrained and real-world oriented problem formulation. First, we employed a peer-reviewed dataset with an equal number of images across all categories [4]. In addition, our dataset contains RGB images captured with a smartphone camera. Furthermore, we use a stratified 5-fold cross-validation to prevent lucky splits in our training and test sets. Without cross-validation, there is no guarantee that a lucky split will not occur, which makes the results less reliable [36]. In contrast to our approach, Islam et al. [14] use a photometric stereo sensor, where four color images are captured for each fabric sample under different illumination conditions [14,37]. Such data acquisition simplifies the prediction task compared to classification from single RGB images. This introduces additional challenges in ensuring that images of the same fabric are strictly separated between the training and testing sets. The absence of cross-validation further increases this probability, as different data splits may include varying numbers of the same images in the training and testing sets. Therefore, we establish a reproducible cross-validated camera-based benchmark using stratified 5-fold cross-validation.

Use-case implications

As we set a new benchmark for visually classifying the cotton content of fabrics, we show that a hybrid approach is applicable for sufficient classification. By not depending on NIR spectrography or other complicated techniques, our approach can be less costly and complex. The usability of only this approach for tasks such as correctly labeling fabrics is, even with this new benchmark, limited, but it could be used to pre-classify the fabrics to speed up other processes. Here, after the fabrics are delivered, our system could be applied for an initial sorting task. The same goes for the application in recycling tasks, where a lot of different fabrics have to be sorted. In this area, it is not essential to classify correctly on the percent, but DL-methods like ours can help to organize the fabrics in similar groups.

Circular economy implications

We set out to help solve issues of the circular economy, as well as applications in the textile industry, and find possible applications for recycling practices. These are important for our world’s sustainable development, as recycling and waste reduction through reuse are key targets for the SDG goal 12, sustainable consumption and production patterns [2]. A possible route for the recycling industry to enhance its efficiency, especially through the application in sorting tasks are benefits of machine-learning approaches [38]. Furthermore, automating the sorting process is a strong hope to solve a key challenge of the industry, as it improves the economic viability of textile waste recycling [39]. As this sorting task could be a key use case of our architecture, we propose a possible approach to enhance the efficiency and economic viability of textile recycling, which helps to come a little closer to reaching the UN SDGs.

Limitations

We set a new benchmark for visual fabric classification; however, our work does come with some limitations. Our model does have a low RMSE and MAE; however, it tends to incorrectly classify neighboring classes. While we achieve good performance metrics, the performance of our model is not enough for high precision tasks, such as determining the cotton content of a fabric to the percent. This must be viewed against the background of the inherent limitations in accuracy that visual methods face compared to spectrographic methods. Visual methods are used because they are quick and easily scalable; however, they measure a proxy for the cotton content as they capture the visual appearance. It is therefore expected that lower performance will be achieved compared to, e.g., spectroscopy, where the chemical composition is directly part of the measurement signal. Since it is conceivable for visual images that identical cotton contents lead to identical visual appearance, zero-error performance is impossible due to this accuracy ceiling of visual methods alone. Thus, a 14.01% RMSE can be seen as a strong achievement within the visual domain, given its practical advantages; however, it offers inherent limitations in achieving the highest performance. Another limitation arises from the fact that we classified based on cotton content. Additional modifications in this vein could also improve the relevance to practical tasks, as fabrics like wool or polyester are common as well. Furthermore, the classes utilized for our model classification came from our dataset. Here, a different division of the classes, for example, dividing them into 5% increments, could be implemented.

Future work

Future research could focus on further analyzing the here occurred misclassification clusters. While the problems of CNNs with multiclass problems and common workarounds are extensively researched, this phenomenon is not. Additionally, the proposed architecture could be applied in different areas for similar tasks to further explore the potential of a DenseNet-Swin combination. Similarly, our modifications to DenseNet121 could be applied in different areas to research possible performance enhancements. To find reasons for the misclassification cluster that occurred in our work, explainable AI tools could be applied. Here, common tools like LIME and SHAP proved performance across different application fields [40]. Additionally, future work could explore the application of preprocessing filters like the Gaussian filter, which has been proven to enhance accuracy in multiple application fields [41–43]. Following this example, Gaussian filtering could be tested in our use case as well. Our approach could further be applied to other fields of research to explore the possible performance gains when applying hybrid architectures with the Swin Transformer V2. For lowering costs and enhancing precision in other application fields of the textile industry, the classification of different textures and weaving types could be an interesting approach.