Saturday, 6 June 2026

Understanding the Paper: Traditional Indian Textiles Classification Using Deep Feature Fusion with Curvelet Transforms

Traditional Indian Textiles Classification Using Deep Feature Fusion with Curvelet Transforms

The paper “Traditional Indian Textiles Classification using Deep Feature Fusion with Curvelet Transforms” presents a machine-learning method for classifying traditional Indian textile and art-form images. The work focuses on seven categories: Batik, Chikankari, Ikat, Kalamkari, Kashida, Madhubani, and Warli.

The central idea of the paper is that traditional textile motifs are visually complex. They contain curves, edges, repeated patterns, ornamental structures, and regional design signatures. A normal deep learning model can learn high-level features from images, but it may not fully capture curved edges and directional singularities. To address this, the authors combine deep features from pre-trained CNN models with Curvelet transform features.

1. Problem Addressed by the Paper

Traditional Indian textiles and art forms have rich visual diversity. Motifs from regions such as Bihar, Andhra Pradesh, Uttar Pradesh, Kashmir, and other craft clusters carry cultural identity. These patterns are not merely decorative; they represent region, technique, community knowledge, and artistic tradition.

However, as mechanization increases and handmade practices decline, there is a need to digitally preserve, organize, classify, and retrieve traditional design images. Manual classification is difficult because many traditional motifs are complex, curved, dense, and visually similar across categories.

The paper addresses this challenge by proposing an image-classification framework that combines two kinds of information:

  • deep features extracted from pre-trained CNN models, and
  • Curvelet transform features that capture curved edges and directional structures.
Core research question: Can traditional Indian textile and art-form images be classified more accurately by combining deep CNN features with Curvelet transform features?

2. Why Traditional Textile Classification Matters

Traditional textile designs are important for cultural preservation, design education, e-commerce cataloging, museum documentation, and digital archives. If textile patterns can be classified automatically, designers and researchers can search large databases more efficiently.

For example, an image of a Kalamkari motif, an Ikat pattern, or a Chikankari embroidery design may be automatically tagged and stored under its correct category. This can help designers retrieve similar designs, compare regional motifs, study design evolution, and validate novelty in new design development.

In the context of online shopping, classification also supports better product discovery. Customers may search by craft type, motif type, region, or visual style. Automated classification can improve tagging accuracy and reduce dependence on manual cataloging.

3. Main Idea of the Proposed Method

The proposed method is based on feature fusion. Instead of depending only on one type of feature, the authors combine hand-crafted transform-based features with deep neural network features.

The broad workflow is as follows:

  1. Collect traditional textile and art-form images.
  2. Resize images according to the CNN architecture.
  3. Optionally apply CLAHE pre-processing.
  4. Extract Curvelet transform features from the image.
  5. Extract deep CNN features using InceptionResNetV2 or VGG16.
  6. Concatenate the Curvelet and CNN features.
  7. Train classifiers such as XGB, Gradient Boosting, and Logistic Regression.
  8. Evaluate performance using precision, recall, F1 score, accuracy, and specificity.

The paper shows that the fusion of Curvelet features and deep CNN features can perform better than using CNN features alone.

4. Dataset Used in the Study

The study uses a dataset of 1046 traditional design images across seven classes. Each image has a resolution of \(1200 \times 1200\). The dataset is divided into training and testing sets using a 90:10 split.

Class Number of Images Visual Nature
Batik 145 Wax-resist textile patterns with organic and decorative forms
Chikankari 155 Embroidery-based motifs, often delicate and floral
Ikat 163 Resist-dyed yarn patterns with blurred geometric edges
Kalamkari 130 Hand-drawn or block-printed narrative and floral motifs
Kashida 141 Embroidery motifs often inspired by nature and regional ornamentation
Madhubani 158 Dense folk-art compositions with line work and symbolic motifs
Warli 154 Tribal art style with human figures, animals, and geometric forms

5. Pre-processing Using CLAHE

The paper uses Contrast Limited Adaptive Histogram Equalization, commonly called CLAHE, as a pre-processing step. CLAHE improves local contrast in images and can help reduce variability caused by lighting, contrast, and image-capture conditions.

In simple terms, CLAHE improves the visibility of local patterns without over-amplifying noise. This is useful in textile images because motifs may contain fine lines, texture differences, and subtle contrast changes.

The images are resized depending on the CNN architecture:

  • \(299 \times 299\) for InceptionResNetV2
  • \(224 \times 224\) for VGG16

The paper observes that pre-processing does not help all models equally. For InceptionResNetV2 combined with Curvelet features, the best result is achieved without pre-processing. For VGG16 combined with Curvelet features, pre-processing improves performance.

6. Curvelet Transform Features

Curvelet transform is one of the key ideas in the paper. Traditional textile motifs often contain curves, bends, ornamental outlines, and directional edges. Wavelets are useful for detecting discontinuities, but they are less effective in representing smooth curves along edges. Curvelets are designed to represent curved singularities more effectively.

A simplified way to understand Curvelet transform is this:

Wavelets are good at point-like discontinuities. Curvelets are better at curve-like and edge-like structures.

The paper uses Fast Discrete Curvelet Transform to obtain Curvelet coefficients from color images. The low-frequency approximate Curvelet coefficients are used as features. The authors test different decomposition scales and use:

  • 4-scale Curvelet features for \(299 \times 299\) images, and
  • 5-scale Curvelet features for \(224 \times 224\) images.

The general feature-extraction idea may be represented as:

\[ I(x,y) \rightarrow C_{s,\theta}(x,y) \]

Here, \(I(x,y)\) represents the input image, \(s\) represents scale, \(\theta\) represents orientation, and \(C_{s,\theta}(x,y)\) represents the Curvelet coefficient at a particular scale and orientation.

The full Curvelet feature vector can be conceptually written as:

\[ F_{curvelet} = [C_{1,\theta_1}, C_{1,\theta_2}, \ldots, C_{s,\theta_k}] \]

This feature vector captures directional and curved structures in the image.

7. Deep Feature Extraction Using CNNs

The paper uses two pre-trained CNN architectures as feature extractors: InceptionResNetV2 and VGG16. These models were originally trained on ImageNet and are reused here through transfer learning.

InceptionResNetV2

InceptionResNetV2 is a deep architecture that combines Inception-style multi-scale convolutional processing with residual connections. It is useful because textile images may contain motifs at different scales: small lines, medium ornamental units, and larger design layouts.

For this model, the input image size is:

\[ 299 \times 299 \]

VGG16

VGG16 is a simpler CNN architecture with stacked convolutional layers. It is easier to compute than InceptionResNetV2 and is often used as a baseline feature extractor in image-classification tasks.

For this model, the input image size is:

\[ 224 \times 224 \]

The deep CNN feature vector may be represented as:

\[ F_{CNN} = f_{CNN}(I) \]

Here, \(f_{CNN}\) represents the feature-extraction function of the pre-trained CNN, and \(I\) is the input image.

8. Feature Fusion Strategy

The main contribution of the paper lies in combining Curvelet features with deep CNN features. This is called feature fusion.

The fusion can be represented as:

\[ F_{fusion} = [F_{CNN}; F_{curvelet}] \]

Here, \(F_{CNN}\) captures high-level learned visual features, while \(F_{curvelet}\) captures curve-based and directional texture features. The semicolon indicates concatenation of the two feature vectors.

This is a useful strategy because the two feature types are complementary. CNNs learn abstract visual patterns from data, while Curvelet transforms explicitly capture curved edges and singularities. Traditional textile designs often depend on both kinds of visual information.

Feature Type What It Captures Why It Helps in Textile Classification
Deep CNN features High-level learned patterns Useful for recognizing overall visual style and motif structure
Curvelet features Curves, edges, orientations, and singularities Useful for traditional motifs with curved ornamental lines
Fused features Combined learned and transform-based representation Improves classification performance by using complementary information

9. Classifiers Used

After feature extraction and fusion, the authors use three classifiers:

  • Gradient Boosting Classifier
  • Extended Gradient Boosting, also called XGB
  • Logistic Regression

Among these, XGB gives the best result when used with InceptionResNetV2 and Curvelet features.

The classification task can be written as:

\[ \hat{y} = g(F_{fusion}) \]

Here, \(F_{fusion}\) is the final feature vector, \(g\) is the classifier, and \(\hat{y}\) is the predicted textile class.

10. Grad-CAM Visualization

The paper also uses Gradient-weighted Class Activation Mapping, or Grad-CAM, to understand which regions of the image influence CNN-based classification. Grad-CAM produces heatmaps that highlight important regions for prediction.

The paper shows Grad-CAM and Guided Grad-CAM examples for Madhubani and Chikankari images. These visualizations help identify the image regions that the model considers important for classification.

In textile research, such visualization is valuable because it can tell us whether the model is looking at meaningful motif areas or irrelevant background regions.

11. Results and Interpretation

The best-performing method is the combination of InceptionResNetV2 and Curvelet features without pre-processing, classified using XGB. This method achieves:

Metric Best Reported Result
Precision 98.24%
Recall 97.15%
F1 Score 97.15%
Accuracy 97.15%
Specificity 99.52%

The result shows that Curvelet features add useful information to deep CNN features. InceptionResNetV2 alone performs strongly, but adding Curvelet features improves the performance further.

The paper also reports that VGG16 paired with Curvelet features improves over VGG16 alone, though it does not reach the performance of InceptionResNetV2 paired with Curvelets. This suggests that the stronger deep feature extractor, when fused with Curvelet information, gives the most effective representation.

Model Combination Reported Accuracy Interpretation
InceptionResNetV2 + Curvelet + XGB 97.15% Best overall result
Pre-processing + Curvelet + InceptionResNetV2 95.24% Strong, but lower than without pre-processing
Pre-processing + Curvelet + VGG16 91.43% Best VGG16-related fusion result
Curvelet + VGG16 88.57% Fusion improves over weaker feature sets but remains below IRV2 fusion
Main result: The fusion of deep features and Curvelet features improves classification of traditional Indian textile images, especially when InceptionResNetV2 features are fused with 4-scale Curvelet features.

12. Relevance for Saree and Textile Research

This paper is highly relevant for saree and textile image classification because many textile identities are based on visual structure, motif vocabulary, and regional design grammar. Saree provenance classification often requires recognizing curved motifs, borders, pallus, repeat structures, and weave-related textures.

A CNN can learn powerful abstract visual features, but textile motifs often contain fine directional details. Curvelet features may help capture curved edges and ornamentation that are especially common in Indian textile designs. Therefore, feature fusion can be a useful direction for saree-origin classification.

For saree classification, this idea can be extended in the following ways:

  • combine CNN features with Curvelet features for motif and border recognition;
  • use Grad-CAM to check whether the model is attending to textile-relevant regions;
  • compare feature fusion with Vision Transformer and graph-based approaches;
  • use retrieval systems to find visually similar sarees or design motifs;
  • build craft-wise digital archives for design preservation and education.
Paper Concept Possible Saree Classification Use
Curvelet transform Capture curved motifs, ornamental borders, and pallu line structures
InceptionResNetV2 features Extract high-level visual style and motif composition
Feature fusion Combine handcrafted texture information with learned deep features
XGB classifier Classify fused features into craft or regional categories
Grad-CAM Interpret whether the model is focusing on meaningful textile regions

13. Limitations and Future Scope

The paper reports promising results, but it also recognizes that further work is needed on larger and more complex datasets. The dataset contains 1046 images across seven classes. For real-world textile classification, larger datasets with more variations in lighting, image angle, fabric folds, motif scale, and product type would be required.

Another limitation is that the study focuses on image-level classification. In actual textile analysis, some categories may require localized understanding of motifs, borders, weave structures, embroidery stitches, and finishing details. Future research may combine classification with object localization, segmentation, and content-based image retrieval.

The authors suggest that the proposed work may be extended toward content-based retrieval of similar designs, novelty validation, and creative AI-based design generation.

14. Simple Summary

This paper proposes a method for classifying traditional Indian textile and art-form images using a combination of Curvelet transform features and deep CNN features. Curvelet features are useful because traditional motifs often contain curved edges and directional structures. CNN features are useful because they capture high-level visual patterns.

The best result comes from combining InceptionResNetV2 features with 4-scale Curvelet features and classifying them using an XGB classifier. This combination achieves 97.15% accuracy on a seven-class dataset containing Batik, Chikankari, Ikat, Kalamkari, Kashida, Madhubani, and Warli.

For textile and saree research, the paper is important because it shows that deep learning can be strengthened by adding transform-based features. This is especially useful in domains where motif curves, edge flow, and regional design details matter.

15. General Disclaimer

This article is an educational explanation of the research paper “Traditional Indian Textiles Classification using Deep Feature Fusion with Curvelet Transforms.” It is intended for conceptual understanding, academic discussion, and research learning. Some technical details have been simplified for readability. Readers interested in exact experimental settings, hyperparameters, dataset preparation, and complete numerical results should refer to the original paper.

```

No comments:

Post a Comment

Understanding the Paper: Drishtikon

DRISHTIKON: A Multimodal Multilingual Benchmark for Indian Cultural Understanding The paper “DRISHTIKON: A Multimodal Multilingual Benchm...