Design and Development of a Handloom Saree Testing System Using Machine Learning
The paper “Design and Development of Handloom Saree Testing System using Machine Learning” presents a practical machine-learning-based system for identifying whether a silk saree has been woven on a handloom or a powerloom. The work is especially important because handloom sarees command higher cultural, emotional, and commercial value, while powerloom sarees can sometimes visually imitate handloom products.
The authors propose a low-cost testing system that scans a saree at high resolution, processes the image, segments the fabric structure using K-means clustering, extracts weave-related features, and finally classifies the saree using a Decision Tree classifier. The system is positioned as an objective alternative to conventional expert-based visual inspection.
Table of Contents
- The Problem Addressed by the Paper
- Handloom vs Powerloom: Why Identification Matters
- Main Idea of the Proposed System
- Hardware Design
- Software Workflow
- Role of K-means Clustering
- Feature Extraction from Weft Irregularity
- Decision Tree Classification
- Performance Metrics
- Experimental Results
- Importance for Textile Testing
- Limitations and Future Scope
- Simple Summary
- General Disclaimer
1. The Problem Addressed by the Paper
Silk sarees are deeply connected with Indian culture and textile heritage. Many weaving clusters in India have their own distinctive saree traditions, and handloom sarees are often valued more highly because they involve skilled manual weaving, softer drape, greater resilience, and visible human craftsmanship.
However, distinguishing handloom from powerloom is not always easy for ordinary customers. Traditionally, this identification is done by experts who observe the weave, texture, irregularities, floats, loose threads, and pallu details. Such methods are subjective and depend heavily on the skill and experience of the examiner.
The paper tries to solve this problem by developing a more objective, machine-assisted testing method. Instead of relying only on human judgment, the proposed system uses high-resolution imaging and machine learning to analyze structural differences in the saree fabric.
2. Handloom vs Powerloom: Why Identification Matters
The paper explains that handloom and powerloom differ in the way the weft is inserted during weaving. In handloom weaving, the shuttle carrying the weft yarn is thrown manually. In powerloom weaving, the pick motion is achieved through machines.
This difference in production method affects the final fabric. Handloom sarees may show slight irregularities because of human involvement. Powerloom sarees, on the other hand, tend to be more uniform and mechanically even. This difference in weave evenness becomes the basis for the proposed machine-learning approach.
| Aspect | Handloom Saree | Powerloom Saree |
|---|---|---|
| Production method | Manual weaving with human-controlled pick motion | Machine-driven weaving with mechanized pick motion |
| Texture | Usually softer and more resilient | Often stiffer and more mechanically uniform |
| Weave appearance | May contain minor irregularities | Often more even and flawless |
| Commercial value | Generally higher because of manual craftsmanship | Generally lower because production is faster |
| Testing challenge | Requires expert visual judgment | Can visually imitate handloom in some cases |
3. Main Idea of the Proposed System
The proposed system is built around one practical insight: the spacing and regularity of weft lines can reveal whether a saree is handloom or powerloom. Since handloom weaving is manually controlled, the spacing between weft insertions may show small irregularities. Since powerloom weaving is machine controlled, the weave structure may show greater regularity.
The system therefore scans a saree sample, detects weft-related structural patterns, measures the variation between consecutive weft lines, and uses these measurements to classify the saree.
The broad pipeline is:
- Scan the saree at high resolution.
- Pre-process the image using blurring and color-space conversion.
- Segment the image using K-means clustering.
- Identify the best cluster representing weft structure.
- Extract features such as spacing variation and frequency.
- Use a Decision Tree classifier to classify the saree as handloom or powerloom.
- Display and generate a report through a graphical user interface.
4. Hardware Design
The hardware setup uses a high-resolution Canon Canoscan LIDE400 scanner, a Raspberry Pi 4 microcontroller board, a monitor, keyboard, mouse, and power adaptor. The scanner captures the saree image at a microscopic level, while the Raspberry Pi performs image processing and classification.
The paper reports that the full standalone system can be built for around $270. If a laptop or desktop is already available, the system cost can come down to around $100, mainly because the scanner becomes the key additional hardware component.
| Component | Purpose | Approximate Cost Reported |
|---|---|---|
| Canoscan scanner | Captures high-resolution saree images | $94 |
| Raspberry Pi 4 | Runs image processing and machine learning pipeline | $97 |
| Monitor, keyboard, mouse | User interaction and display | $75 |
| Power adaptor | Power supply for electronics | $4 |
| Total | Standalone testing system | $270 |
5. Software Workflow
The software system is developed using Python and common computer vision and machine learning libraries such as OpenCV, sklearn, imutils, and numpy. A graphical user interface is also developed to make the system easier to use.
The image processing workflow begins with a high-resolution scan of a small saree region. The scanned image is then blurred using Gaussian blurring to smooth edges and reduce noise. The paper mentions a Gaussian window size of \(21 \times 21\) pixels.
The blurred image is converted from RGB to HSV color space. The authors found that HSV produced better segmentation results than RGB for this application.
The image is then segmented using K-means clustering with \(K = 4\). After segmentation, the cluster with the least deviation is selected as the best cluster for further feature extraction.
6. Role of K-means Clustering
K-means clustering is used as an unsupervised image segmentation method. Since the scanned saree image contains different visual elements such as yarn regions, background, shadows, and weave structures, segmentation helps isolate useful structural patterns.
In K-means clustering, pixels are grouped into \(K\) clusters based on similarity. The objective is to minimize the distance between each point and its assigned cluster center.
A simplified form of the K-means objective is:
\[ J = \sum_{i=1}^{K} \sum_{x \in C_i} \|x - \mu_i\|^2 \]
Here, \(C_i\) represents the \(i^{th}\) cluster, \(x\) represents a data point or pixel feature, and \(\mu_i\) represents the centroid of the cluster.
In the paper, \(K = 4\) is reported as the optimum value based on evaluation. After clustering, each cluster is examined, and the cluster with the least deviation is considered the best label for extracting weave-related features.
7. Feature Extraction from Weft Irregularity
After segmentation, the selected cluster is sliced into six pieces and converted into a binary image. Slicing is used to reduce errors caused by slight misalignment when the saree is placed on the scanner.
The system then detects weft lines and calculates spacing-related features. The paper uses \(\Delta\) to represent the distance between two consecutive weft lines. It then calculates \(\Delta\Delta\), which represents the difference between two consecutive \(\Delta\) values.
The important feature idea can be written as:
\[ \Delta_i = d_{i+1} - d_i \]
where \(d_i\) and \(d_{i+1}\) are positions of consecutive weft lines.
The variation between consecutive spacing values is:
\[ \Delta\Delta_i = \Delta_{i+1} - \Delta_i \]
The system then extracts statistical features such as:
- distance between consecutive weft lines, \(\Delta\)
- difference between consecutive spacing values, \(\Delta\Delta\)
- frequency of occurrence
- mean of \(\Delta\Delta\)
- standard deviation of \(\Delta\Delta\)
A histogram is generated from the extracted features. The x-axis represents \(\Delta\Delta\) values, and the y-axis represents their frequency of occurrence.
8. Decision Tree Classification
The final classification is performed using Decision Tree classifiers. A Decision Tree is a supervised machine learning model that uses a tree-like structure of decision rules. It starts from a root node, passes through decision nodes, and ends at leaf nodes that represent final class labels.
In this paper, two Decision Tree classifier models are developed:
| Model | Feature Used | Purpose |
|---|---|---|
| DTC Model 1 | Optimized \(\Delta\Delta\) feature | Classifies saree based on variation in weft spacing |
| DTC Model 2 | Optimized frequency feature | Classifies saree based on frequency pattern of weave variation |
The paper reports that Model 2, which uses the optimized frequency feature, performs better than Model 1.
9. Performance Metrics
The proposed system is evaluated using common classification metrics: recall, precision, F1 score, and accuracy.
Recall measures how many actual positive cases were correctly identified:
\[ Recall = \frac{TP}{TP + FN} \]
Precision measures the quality of positive predictions:
\[ Precision = \frac{TP}{TP + FP} \]
F1 score combines precision and recall:
\[ F1 = 2 \times \frac{Precision \times Recall}{Precision + Recall} \]
Accuracy measures the overall correctness of the model:
\[ Accuracy = \frac{Correct\ Predictions}{Total\ Predictions} \]
Here, \(TP\) means true positives, \(FP\) means false positives, and \(FN\) means false negatives.
10. Experimental Results
The system was tested at ten silk saree showrooms in and around Bangalore, with nearly 150 sarees. The authors observed that powerloom sarees showed higher frequency patterns than handloom sarees because powerloom fabrics tend to have more even texture.
The reported performance comparison is:
| Metric | DTC Model 1 | DTC Model 2 |
|---|---|---|
| Recall | 0.86 | 0.92 |
| F1 Score | 0.86 | 0.88 |
| Precision | 0.86 | 0.84 |
| Accuracy | 0.81 | 0.83 |
The best recall value reported is \(0.92\), achieved by the Decision Tree model using the optimized frequency feature.
The paper also compares processing time on Raspberry Pi and laptop hardware. The total processing time is reported as approximately 171 seconds on Raspberry Pi and 81 seconds on a laptop. The authors suggest that on a higher-end system, the overall processing time could potentially be reduced to less than a minute.
| Hardware | Total Processing Time |
|---|---|
| Raspberry Pi 4 | Approximately 171 seconds |
| Laptop | Approximately 81 seconds |
11. Importance for Textile Testing
The paper is important because it moves saree identification from subjective judgment toward objective textile testing. In the traditional market, authenticity is often dependent on expert opinion. But expert opinion can vary, and customers may not have access to trained examiners.
A machine-learning-based testing system could support retailers, testing laboratories, certification agencies, and consumers by providing a more standardized method. The paper also mentions the possibility of recognition by textile testing laboratories and eventual validation by standards bodies.
For handloom clusters, such systems may help protect authentic products from imitation. For customers, they may increase confidence while purchasing high-value sarees. For researchers, the work shows how image processing and simple machine learning can be applied to a culturally important textile problem.
12. Limitations and Future Scope
The proposed system is practical and promising, but it also has some limitations. The dataset used in the study is relatively small, with nearly 150 sarees. A larger and more diverse dataset across different silk saree clusters would make the model more reliable.
The method focuses mainly on weave regularity and weft spacing. However, saree authenticity can also involve yarn type, zari quality, motif construction, pallu technique, border joining, loom setting, finishing, and regional craft grammar. Future systems may combine microscopic weave analysis with broader image-based and material-based features.
Another limitation is that the system requires careful scanning. If the fabric is tilted, folded, shiny, or poorly placed, the captured image may affect segmentation and feature extraction. The paper tries to handle this through slicing, but real-world deployment would still require careful operating procedure.
| Limitation | Possible Improvement |
|---|---|
| Small dataset | Test on thousands of sarees from multiple clusters |
| Focus on weft regularity | Add features related to yarn, weave, motif, border, and pallu |
| Scanner placement sensitivity | Develop standardized scanning fixtures and alignment correction |
| Simple classifier | Compare with Random Forest, SVM, CNN, and hybrid models |
| Limited product diversity | Evaluate across silk, cotton, blended, jacquard, and regional saree types |
13. Simple Summary
This paper proposes a practical system for identifying whether a silk saree is handloom or powerloom. The system scans a saree, processes the image, segments the weave structure using K-means clustering, extracts spacing and frequency features from the weft pattern, and classifies the saree using a Decision Tree model.
The main insight is that handloom sarees tend to show more human irregularity, while powerloom sarees tend to show more mechanical regularity. By converting these differences into measurable image features, the system attempts to make saree testing more objective.
The reported results are encouraging: the best model achieves a recall of \(0.92\), the laptop-based system gives a response time of around 81 seconds, and the system can be built at relatively low cost. For the handloom sector, this kind of system can support authenticity testing, consumer confidence, and protection of craft value.
14. General Disclaimer
This article is an educational explanation of the research paper “Design and Development of Handloom Saree Testing System using Machine Learning.” It is meant for learning, discussion, and conceptual understanding. The explanation simplifies some technical details for readability. For complete methodology, experimental details, hardware configuration, and exact validation procedure, readers should refer to the original paper.
No comments:
Post a Comment