www.rsisinternational.org

Page 788

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

Deep Learning Approaches for Automatic Recognition of Textile

Weave Structures

*Pugazhenthi T.

& Pravin P Chavan.

National Institute of Fashion Technology, Textile Design Department, Kannur, Kerala, India

National Institute of Fashion Technology, Knitwear Design Department, Kannur, Kerala, India

DOI:

https://doi.org/10.51583/IJLTEMAS.2026.150600060

Received: 14 June 2026; Accepted: 19 June 2026; Published: 06 July 2026

ABSTRACT

The Fabric weave structure is an essential component in the design and manufacture of premium fabric. The

structure or pattern used in the fabric's weaving significantly impacts its overall appearance, texture, durability,

and drape. Traditional methods of Weave pattern recognition are relied heavily on manual visual inspection,

which is time-consuming, prone to human error, struggles with complex patterns and sensitive to lighting

conditions Therefore, an automated system is necessary for the classification of woven cloth to improve

production efficiency. In this paper we proposed a deep learning model particularly Transfer Learning model

that employs data augmentation and transfer learning methods for the classification and identification of woven

textiles and compared with other models. The model uses a MobileNet-V2 to autonomously extract and

categorize fabric texture features in an end-to-end fashion. The model is trained using a custom dataset with

various weave types, including plain, twill and satin. The experimental findings indicate that the suggested model

is resilient and attains cutting-edge precision. We compared our findings with other baseline methods,

demonstrating that the suggested technique attained superior accuracy when accounting for rotational

orientations in fabric and appropriate illumination conditions.

Keywords: Deep learning, MobileNet-V2, Pattern recognition, ResNet-50, Transfer learning, Woven fabric

INTRODUCTION

Pattern recognition in weaving is an essential function within the textile industry, influencing fabric quality,

design, and manufacturing efficiency. It is regarded as a crucial process prior to fabric production. The

importance of classifying textile weave patterns stems from its considerable influence on fabric aesthetics,

textural evaluation, and design modification. Conventional methods for weave categorization rely heavily on

manual visual evaluation, which is laborious and prone to human error. This approach may lead to discrepancies,

especially when dealing with complex patterns (Das et al., 2022). Some traditional methods are very dependent

on controlled lighting conditions.

Fluctuations in lighting can markedly influence the precision of texture analysis and pattern identification

(Akram et al., 2024; Kuo et al., 2010). Certain conventional techniques necessitate the physical extraction of

yarns from the fabric to examine the weave pattern, which may compromise the fabric and is inappropriate for

non-destructive testing (Sabuncu et al., 2016; Kuo et al., 2016). Conventional techniques frequently encounter

challenges regarding precision and consistency, especially when addressing complex or non-repetitive weaving

patterns. These approaches may inadequately capture the nuanced differences in fabric textures, resulting in

misclassification (Kuo et al., 2010; Meng et al., 2020). To enhance the efficiency of textile production, numerous

researchers have dedicated themselves to employing image processing and artificial intelligence technologies

for the automatic identification of weave patterns. Over the years, various image processing techniques have

been developed to automate this process, reducing reliance on manual inspection and improving accuracy.

www.rsisinternational.org

Page 789

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

METHODS

CONVOLUTIONAL NEURAL NETWORKS ARCHITECTURE (CNN)

CNNs have emerged as a powerful tool for fabric weave pattern recognition. These deep learning models excel

in identifying patterns in textile images due to their ability to learn hierarchical features directly from data.

Several studies have utilized CNNs for this purpose. A basic CNN model was used to classify three textile weave

patterns (plain, twill, and satin) with an average accuracy of 72%. However, the model performed differently

across patterns, achieving 100% accuracy for plain weave but only 50% for satin weave (Pasha et al., 2024).

VGG16 Model

A modified VGG16 model with additional pooling layers was proposed to improve accuracy and reduce

computational requirements. This model achieved 90% accuracy and an F1-score ranging from 0.8 to 1 (Akram

et al., 2024).

Resnet Architecture

Rauf et al. (2022) used ResNet, a pre-trained CNN architecture, in combination with texture features extracted

via the gray-level co-occurrence matrix (GLCM) and Gabor wavelets, achieving a high accuracy of

96.15%.CNNs have proven to be highly effective for texture-based classification tasks, though their performance

can vary depending on the complexity of the weave patterns and the quality of the training data.

Transfer Learning

The choice of pre-trained model architecture significantly impacts the accuracy of textile weave pattern

classification using transfer learning, as evidenced by various studies. DenseNet121, for instance, has been

shown to achieve high accuracy in classifying woven fabric patterns, with a reported accuracy of 92.58% and a

loss of 29.62%, outperforming other model such as MobileNetV2 in specific applications (Maur et al., 2023).

The model's efficiency is attributed to its compact size and effective feature propagation, making it suitable for

mobile applications.

Proposed Model

We designed our deep learning model as a pipeline based approach which includes basic CNN and MobileNet-

V2 model. The pipeline was structured to have a certain number of phases, with the first stage including the

fabric images and the last stage encompassing the model categorization. The output of one step served as the

input for the subsequent stage. The planned pipeline is seen in Figure 1 and delineated as follows:

Figure 1. Proposed pipeline approach.

www.rsisinternational.org

Page 790

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

Data Collection and Preprocessing

The woven fabric images were collected into a database and stored as 2 dataset namely training and testing.

The data set needed appropriate conversions, image resize and preprocessing for training and testing the data.

As the number of images is still a relatively small, it is important to expand the dataset and we applied some

augmentation methods to expand the data set to help the model have a good generalization, achieve the better

identification.

Model Generation and Training

A learning method that processes input data that mapped to characteristics corresponding to the goal and predict

the output class is referred to as a model. We used the MobileNet-V2 architecture for our model. During the

training process, the initial convolutional layers were left alone, and only the custom layer linked to the base

network was trained.

Model Evaluation

Our proposed model is evaluated across a range of performance metrics including accuracy, balanced accuracy,

precision, recall, and F1-score.

Dataset

The model was trained and tested on a Digital Microscope with Camera Endoscope 8-LED Light Magnifying

Glass Magnifier (1000× zoom) of woven fabric texture images. Total 1182 images have been taken for three

kinds of fabrics which includes 532 images for Plain cloth, 504 images for Twill and 146 images for Satin. Of

total 1182 images, 20% images are used as validation images and remaining 80% images are as train images.

where 80% and 20% from 1182 images means 946 images are as training image and 236 images from total

1182 are as validation images.

Preprocessing

During preprocessing, each image is first converted from RGB (color) to grayscale so that the model learns the

fabric weave pattern instead of the color. The grayscale image is then converted back to a 3-channel image

because MobileNet requires a 3-channel input. Finally, the pixel values are normalized to the range -1 to 1,

which helps the model train more effectively. The same preprocessing steps are applied to both the training and

validation datasets to ensure consistency.

According to Shorten and Khoshgoftaar (2019), data augmentation is a group of methods that make training

datasets larger and more diverse, thereby improving deep learning model performance. Hussain et al. (2020)

reported that the limitation of a small training dataset can be addressed through data augmentation techniques.

Data augmentation applies various transformations, including zoom in, zoom out, scaling, skew, flip, and

intensities to the whole dataset to obtain a diverse set of images thereby enlarging the dataset

Deep learning models perform particularly well when trained on large datasets. Data augmentation increases

the size of the training dataset, enabling the model to learn more effectively. Data warping techniques generate

additional samples by applying various transformations to the original data, thereby enhancing the diversity of

the training set (Wong et al., 2016).

This process helps reduce overfitting and improves the model's generalization capability without altering the

underlying data characteristics. In this study, several augmentation techniques were applied to the images,

including horizontal and vertical flipping, shifting, rotation, zooming, shearing, and brightness adjustment.

www.rsisinternational.org

Page 791

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

RESULTS AND DISCUSSION

The MobileNet transfer learning framework, that was discussed about earlier, was used to build our model for

classifying the images. The proposed pre-trained transfer learning model, MobileNet-V2, classifies woven

fabric images into three categories. The quantity of training and testing images in the woven fabric dataset, after

augmentation, has been elevated to 1182. This article utilizes 80% of the training images (946) for model

training, while the remaining 20% (236) is designated as a validation subset to assess the model during data

training. The deep learning model's performance was evaluated using the test set.

Experimental Framework

Initially, all woven fabric images were resized to 224 × 224 pixels. MobileNetV2 was selected as the base

model because of its lightweight architecture and reduced number of parameters. The pretrained MobileNetV2

model was trained using the woven fabric dataset presented in this study. In the first modification, the last two

dense layers of the original MobileNetV2 architecture were removed and replaced with customized dense

layers. Transfer learning was employed to train the newly added layers while keeping the weights of the

remaining layers frozen. Freezing these layers helped accelerate convergence and prevented gradient explosion

during training. In the second modification, to reduce overfitting, the final dense layer was retrained using

additional data, and a dropout layer was inserted between the last two dense layers (Mahanta et al., 2024).

A global average pooling layer followed by batch normalization, two fully connected layers, and two dropout

layers were stacked on the base network. The dense layers contained 128 and 64 neurons, respectively. Each

fully connected layer employed a ReLU activation function, while dropout was used to prevent the network

from relying excessively on specific neurons and to encourage the learning of more robust and generalized

features (Mahanta et al., 2024).

The training efficiency and stability of the pre-trained model were further enhanced through the incorporation

of batch normalization layers (Hidajat et al., 2025). In addition, the combination of global average pooling and

dropout layers helped mitigate overfitting. By randomly deactivating redundant neurons during training, the

dropout layers improved the model's generalization performance. Overfitting is a common challenge in deep

learning architectures, often resulting in poor performance on unseen test data. Finally, a Softmax layer was

employed as the output layer to classify woven fabric images into three categories. A high-level architecture of

the customized deep learning model is presented in Figure 2.

Figure 2. Architecture of the Proposed Model

After feature selection, classification was performed for predicted class against actual class. The model under

consideration had an optimizer (the optimizer to train the parameters, here Adam) and a loss function for

stochastic optimization. The learning rate was initialized to 0.0051. The dropout proportions for the two

dropout layers were set to 0.550. The batch size was set to 32.

www.rsisinternational.org

Page 792

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

Result

The model was trained and tested on a Digital Microscope with Camera Endoscope 8-LED Light Magnifying

Glass Magnifier (1000× zoom) for extraction of woven fabric texture images The model was implemented in

Python 3.9 with Keras serving as a frontend and a Tensorflow a backend.

Evaluation Metrics

Accuracy is the primary metric used to measure the performance of a classification model. It represents the

percentage of correctly classified instances out of the total instances.

Higher accuracy in a model indicates reliability, while large accuracy in an imbalanced dataset suggests bias

favoring the class with more images.. The confusion matrix was calculated to visualize the model's predictions

on test data and understand the total number of correctly and misclassified images, with rows and presented in

Figure 3.

Figure 3. Architecture of the Proposed Model

Table 1. Precision, recall, and F1-score

The precision, recall, and F1-score of the proposed model were averaged over the test set. Performance analysis

of these values in terms of each class is shown in Table 1. Average precision, recall and F1-score were 94.66

Name of the weave

Precision

Recall

F1-score

Support

Plain

0.92

0.96

0.94

115

Satin

0.95

0.97

Twill

0.96

0.89

0.93

www.rsisinternational.org

Page 793

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

DISCUSSION

In previous works, researchers proposed methods that employed traditional machine learning models with

hand-crafted features, in a way being more laborious and time-consuming.

Akram et al. (2024) employed a modified VGG-16 algorithm to extract image features and classify various kinds

of woven fabrics. Their proposed method achieved an accuracy of 90%. Xiao et al. (2018) proposed another

method based on Transform Invariant Low-rank Textures (TILT) and Histogram of Oriented Gradients (HOG),

achieving an accuracy of 94.57% in identifying woven fabric patterns. However, these methodologies were

developed using datasets containing a limited number of images. The authors did not account for physical factors

such as yarn diameter variation, fabric deformation, fabric rotation during image acquisition, uneven

illumination, and other real-world imaging conditions.

With our proposed model, we were able to achieve validation accuracies far superior to random guessing,

providing us with confidence in its effectiveness for our classification task. Furthermore, the data augmentation

techniques increased the diversity of the available data, and as a result, improved the overall performance of the

model. The computational efficiency of this model also made it ideal for deployment in real-world scenarios, as

it minimized processing time and resource usage. Overall, our decision to use our own custom model proved to

be the right one, as it outperformed the pretrained architectures in every aspect important to our project's success.

A MobileNet-V2 architecture (MobileNet-V2)-based pretrained convolutional neural network (CNN) model

was employed as a deep model . The woven fabric images that we had developed were utilized by the proposed

model to execute end-to-end fabric feature extraction and classification.

CONCLUSIONS

In this paper, we proposed a customized deep learning model for the recognition and classification of woven

fabrics. The MobileNet-V2 model clearly outperforms the CNN in the categorization and identification of

fundamental fabric images, including plain, twill and satin. A survey of the current literature indicates that

many authors have contributed to the case study of deep learning in fabric categorization. The primary

justification for endorsing convolutional neural networks (CNN) is the inadequate assessment of fabric by

traditional manual visual techniques, which are inefficient and unlikely to be sustainable in the long term within

industrial and textile sectors. The MobileNet-V2 method effectively reduces computing costs by extracting

essential information throughout the training phase following the addition of additional layers. The experimental

findings indicate that the MobileNet-V2 algorithm outperforms others in accuracy, speed, initial learning rate,

and dropout rate.

www.rsisinternational.org

Page 794

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

REFERENCES

1. Das, S., Das, S. K., Shanmugaraja, K., & Shanmugaraja, K. (2022). Application of artificial neural

network in determining the fabric weave pattern. Zaštita Materijala.

https://doi.org/10.5937/zasmat2203291d

2. Akram, N., Butt, R. A., & Qureshi, M. A. (2024). Modification of a convolutional neural network for

the weave pattern classification. Mehran University Research Journal of Engineering and Technology.

https://doi.org/10.22581/muet1982.2998

3. Kuo, C.-F. J., et al. (2010). Application of computer vision in the automatic identification and

classification of woven fabric weave patterns. Textile Research Journal.

https://doi.org/10.1177/0040517510373630

4. Sabuncu, M., Sabuncu, M., Özdemir, H., & Özdemir, H. (2016). Recognition of fabric weave patterns

using optical coherence tomography. Journal of The Textile Institute.

https://doi.org/10.1080/00405000.2015.1114791

5. Kuo, C.-F. J., Shih, C.-Y., Huang, C.-C., Su, T.-L., & Liao, I.-C. (2016). A novel image processing

technology for recognizing the weave of fabrics. Textile Research Journal, 86(3), 288–301.

https://doi.org/10.1177/0040517514549982

6. Meng, S., et al. (2020). A multi-task and multi-scale convolutional neural network for automatic

recognition of woven fabric pattern. Journal of Intelligent Manufacturing.

https://doi.org/10.1007/s10845-020-01607-9

7. Pasha, Q., Akram, N., Rafique, A., Akber, A., & Abbasi, M. (2024). An automated model for fabric

weave pattern recognition and classification using deep learning based image processing. In 2024 Global

Conference on Wireless and Optical Technologies (GCWOT). IEEE.

https://doi.org/10.1109/GCWOT63882.2024. 10805736

8. Rauf, M. A., et al. (2022). Fabric weave pattern recognition and classification by machine learning. In

2022 2nd International Conference of Smart Systems and Emerging Technologies (SMARTTECH).

IEEE. https://doi.org/10.1109/SMARTTECH54121.2022.00026

9. Maur, Y. E. H., Santoso, A., & Pranowo. (2023). Fine tuned DenseNet121 to classify NTT weaving

motifs on mobile application. International Journal on Advanced Science, Engineering and Information

Technology. https://doi.org/10.18517/ijaseit.13.6.18314

10. Shorten, C., & Khoshgoftaar, T. M. (2019). A survey on image data augmentation for deep learning.

Journal of Big Data, 6(1), 60. https://doi.org/10.1186/s40537-019-0197-0

11. Hussain, M. A. I., Khan, B., Wang, Z., & Ding, S. (2020). Woven fabric pattern recognition and

classification based on deep convolutional neural networks. Electronics, 9(6), 1048.

https://doi.org/10.3390/electronics9061048

12. Wong, S. C., Gatt, A., Stamatescu, V., & McDonnell, M. D. (2016). Understanding data augmentation

for classification: When to warp? In 2016 International Conference on Digital Image Computing:

Techniques and Applications (DICTA) (pp. 1–6). IEEE. https://doi.org/10.1109/DICTA.2016.7797091

13. Mahanta, L. B., Mahanta, D. R., Rahman, T., & Chakraborty, C. (2024). Handloomed fabrics

recognition with deep learning. Scientific Reports, 14(1), 7974. https://doi.org/10.1038/s41598-024-

58750-z

14. Hidajat, M. S., Wibowo, D. A., & Mintorini, E. (2025). Performance comparison of machine learning

algorithms for ikat weaving classification. KINETIK. https://doi.org/10.22219/kinetik.v10i1.2059

15. Xiao, Z., et al. (2018). Automatic recognition of woven fabric pattern based on TILT. Mathematical

Problems in Engineering, 2018, 1–12. https://doi.org/10.1155/2018/9707104