Understading the Paper: Automatic Classification and Color Changing of Saree Components Using Deep Learning Techniques

Automatic Classification and Color Changing of Saree Components Using Deep Learning Techniques

The paper “Automatic Classification and Color Changing of Saree Components using Deep Learning Techniques” presents a practical deep-learning tool for automatically detecting saree components and changing their colors independently. The work focuses on a very real problem in the saree retail and design industry: changing the color of a saree body or border normally requires manual editing in software such as Photoshop, which is time-consuming and requires technical skill.

The authors propose an automated pipeline that identifies whether an image contains a human wearing a saree or only a saree, removes the background where needed, segments saree components such as body and border, and then modifies the color of selected regions using the HSV color space. The method combines Faster R-CNN, MODNet, Mask R-CNN, and HSV-based hue manipulation.

Table of Contents

1. Problem Addressed by the Paper
2. Why Saree Component Segmentation Matters
3. Main Idea of the Proposed Method
4. Overall Architecture
5. Human Detection Using Faster R-CNN
6. Background Removal Using MODNet
7. Saree Body and Border Segmentation Using Mask R-CNN
8. Color Changing Using HSV Color Space
9. Clothing Mask Equation
10. Results and Performance Evaluation
11. Relevance for Saree Retail and Textile Design
12. Limitations and Future Scope
13. Simple Summary
14. General Disclaimer

1. Problem Addressed by the Paper

Sarees have multiple visual components. The body, border, pallu, motifs, zari areas, and background all contribute to the final appearance of the garment. In retail and manufacturing, designers often need to visualize the same saree in different color combinations before producing or displaying it. Traditionally, this requires manual selection and editing of regions in image-editing software.

The paper argues that this process is labor-intensive and expensive. Manual editing also requires technical expertise, and it becomes inefficient when retailers need to experiment with many color combinations across many saree images.

The research therefore proposes an automated system that can identify saree components and change their colors independently. The focus is mainly on the saree body and border.

Core problem: Can deep learning automatically identify saree body and border regions and allow independent color modification without manual selection?

2. Why Saree Component Segmentation Matters

Saree design is not only about the overall color of the garment. The border and body may have different colors, different motifs, different zari details, and different visual importance. A small change in border color can completely alter the visual character of the saree.

For retailers, automatic color modification can help customers visualize color options quickly. For manufacturers, it can reduce sampling cost before production. For e-commerce, it can create product variants faster. For designers, it can help explore color combinations without repeatedly editing images manually.

Application	How Automatic Saree Component Editing Helps
Retail visualization	Customers can see the same saree in multiple color combinations.
Design development	Designers can test body-border color combinations quickly.
E-commerce cataloging	Online platforms can generate visual variants more efficiently.
Manufacturing planning	Color options can be evaluated before physical sampling.
Digital textile tools	Supports automation in traditional apparel visualization.

3. Main Idea of the Proposed Method

The proposed method combines multiple deep-learning and image-processing steps. It does not rely on a single model. Instead, it coordinates different models for different tasks.

The broad workflow is:

1. Classify whether the input image contains a human wearing a saree or only a saree.
2. If a human is present, remove the background using MODNet.
3. Use Mask R-CNN to detect and segment the saree body, saree border, and human parts such as hands and face.
4. If the image contains only a saree, use another Mask R-CNN model to detect body and border regions.
5. Convert the selected region into HSV color space.
6. Find dominant hue values using a hue histogram.
7. Change the hue value of selected pixels while preserving saturation and value.

This allows color modification while retaining the texture, contrast, and brightness of the original saree image.

4. Overall Architecture

The architecture has three main phases:

1. Human detection and background removal
2. Body and border segmentation
3. HSV-based color changing

The paper’s architecture diagram shows a pipeline where Faster R-CNN first classifies whether a human is present. MODNet is used for background removal. Mask R-CNN models then segment saree regions. Finally, the HSV color-changing module modifies selected parts.

Stage	Technique Used	Purpose
Human detection	Faster R-CNN	Determines whether the image contains a human wearing a saree.
Background removal	MODNet	Removes unwanted background noise.
Component segmentation	Mask R-CNN	Segments saree body, border, hands, and face where applicable.
Color modification	HSV color space	Changes selected hue values while preserving contrast and saturation.

5. Human Detection Using Faster R-CNN

The first step is to classify the input image based on whether it contains a human. This is important because images with humans need additional processing. For example, hands, face, and background should not be recolored along with the saree.

The paper uses a Faster R-CNN model trained with more than 1500 images. It reports accuracy above 96% and classification time below 90 milliseconds. This classification determines which processing route the image should follow.

Conceptually, the classification can be represented as:

\[ c = f_{\theta}(I) \]

Here, \(I\) is the input image, \(f_{\theta}\) is the trained Faster R-CNN model, and \(c\) indicates whether a human is present or absent.

6. Background Removal Using MODNet

For images containing a human, background removal is performed using MODNet. Background removal helps isolate the saree and human figure from irrelevant visual noise. This is important because background pixels may introduce additional hues into the image and interfere with color-change accuracy.

MODNet is used because it is lightweight and suitable for real-time or near-real-time background removal. The paper contrasts MODNet with earlier methods such as GrabCut and trimap-free matting approaches, noting that MODNet offers better generalization and efficiency.

Why background removal matters: If background colors remain in the image, the hue histogram may detect them as dominant colors, causing incorrect color modification.

7. Saree Body and Border Segmentation Using Mask R-CNN

The central segmentation task is handled using custom-trained Mask R-CNN models. Mask R-CNN is suitable because it produces pixel-level masks for different object regions. In this study, those regions are saree body, saree border, and, in human images, hands and face.

The paper uses two separate Mask R-CNN models:

Model	Input Type	Classes Segmented
Mask R-CNN Model 1	Images containing humans wearing sarees	Saree body, saree border, hands, and face
Mask R-CNN Model 2	Images of sarees without humans	Saree body and saree border

The models are trained with 500 images each. The images are manually annotated using polygonal regions created with MakeSense AI. The paper reports training with a learning rate of \(0.001\), image size of \(512 \times 512\), batch size of 4, and 15 epochs.

This region-based segmentation is important because sarees have structurally distinct parts. The body and border must be separated before independent recoloring is possible.

8. Color Changing Using HSV Color Space

After identifying the saree body or border region, the system changes color using HSV color space. HSV separates color into three components:

• Hue: the dominant color
• Saturation: the purity or intensity of the color
• Value: the brightness

HSV is useful because the model can modify hue while preserving saturation and value. This means the saree’s original texture, contrast, and brightness are retained even after color change.

In contrast, RGB mixes color, brightness, and intensity across red, green, and blue channels. This makes controlled color replacement more difficult.

The image is converted from RGB to HSV:

\[ I_{RGB} \rightarrow I_{HSV} \]

Then a hue histogram is generated for the segmented region. The dominant hue is identified from the peak of the histogram. Pixels within a tolerance range around this dominant hue are modified to the desired hue.

9. Clothing Mask Equation

The paper defines a mask to identify pixels that belong to the selected saree region and fall within a hue tolerance range. Conceptually, the mask can be written as:

\[ M(i,j) = \begin{cases} 1, & p_{max} - v \leq h(i,j) \leq p_{max} + v \\ 0, & \text{otherwise} \end{cases} \quad \forall (i,j) \in S_{mask} \]

Here, \(S_{mask}\) is the set of pixels detected by Mask R-CNN, \(p_{max}\) is the mode or dominant hue value of the selected pixels, \(v\) is the variance or tolerance range, and \(h(i,j)\) is the hue value of the pixel at coordinate \((i,j)\).

This equation means that only pixels belonging to the detected saree component and having hue values close to the dominant hue are selected for color change.

10. Results and Performance Evaluation

The system was evaluated on a test set of 50 saree images. These images were manually annotated using MakeSense AI to create ground-truth masks for saree body and border. The proposed method’s pixel-level predictions were compared against these annotations.

The paper compares the proposed method with BiSeNetV2 and DeepLabV3+. For saree body segmentation, the proposed method achieves the highest reported accuracy of \(93.01\%\). For saree border segmentation, it achieves \(89.23\%\) accuracy.

Saree Body Segmentation Results

Method	Accuracy	Precision	Recall	F1 Score
BiSeNetV2	83.98	85.32	82.34	83.80
DeepLabV3+	87.45	84.65	86.87	85.74
Proposed methodology	93.01	87.87	86.85	87.36

Saree Border Segmentation Results

Method	Accuracy	Precision	Recall	F1 Score
BiSeNetV2	81.21	80.83	78.55	79.67
DeepLabV3+	86.93	82.57	84.65	83.59
Proposed methodology	89.23	83.41	84.08	83.74

The paper also reports that the proposed algorithm can perform the color-changing task within approximately nine seconds with minimum user effort. This is significant because manual selection and recoloring using image-editing software can be time-consuming.

Main result: The combined use of MODNet, Faster R-CNN, Mask R-CNN, and HSV hue manipulation enables automatic saree component recoloring while preserving visual details.

11. Relevance for Saree Retail and Textile Design

This paper is important for saree retail because it directly addresses a commercial workflow. Retailers often need to show the same saree in multiple color options. Instead of manufacturing every variant or manually editing images, an automated tool can generate realistic visual alternatives.

For textile designers, the method can help evaluate color combinations before production. For e-commerce platforms, it can enable faster catalog generation and more interactive product visualization. For customers, it may improve the buying experience by allowing them to see how different body-border combinations may look.

Paper Concept	Possible Saree Industry Use
Body-border segmentation	Independent recoloring of saree body and border.
Background removal	Cleaner product images for catalog and design visualization.
HSV color manipulation	Change color while retaining folds, texture, and contrast.
Mask R-CNN segmentation	Precise selection of saree parts without manual masking.
Automated pipeline	Reduces editing time for retailers and designers.

12. Limitations and Future Scope

The paper identifies important limitations. The method performs less effectively when images contain unconventional postures. This is understandable because saree draping can create complex folds, occlusions, and body poses.

Another limitation is the difficulty of handling iridescent sarees, especially silk sarees. Silk can show many color variations due to light reflection, angle, weave structure, and yarn luster. When such sarees are recolored, some details may be lost because the image contains a wide range of hue values.

The authors suggest that training on a larger and more diverse dataset can improve performance. They also mention that the method can be extended beyond color alteration to pattern transfer or design modification. The approach may also be applied to other garments such as shirts, trousers, kurtas, blouses, pants, and kimonos.

Limitation	Possible Improvement
Unconventional poses	Train on more diverse draping and posture datasets.
Iridescent silk sarees	Use material-aware color models and illumination correction.
Limited test set	Evaluate on larger datasets across saree types, lighting, and backgrounds.
Color-only modification	Extend to motif, border, and pattern modification.
Body-border focus	Extend segmentation to pallu, blouse, motifs, zari, and fall areas.

13. Simple Summary

This paper proposes an automated method for changing the colors of different saree components. The system first detects whether the image contains a human wearing a saree or only a saree. It removes background where required, segments the saree body and border using Mask R-CNN, and then changes the color of selected regions using HSV hue manipulation.

The key strength of the method is that it allows body and border colors to be changed separately while preserving texture, folds, contrast, and brightness. The reported segmentation accuracy is \(93.01\%\) for saree body and \(89.23\%\) for saree border.

For the saree industry, this work is valuable because it can reduce manual photo-editing effort and help retailers, designers, and customers visualize multiple color options before production or purchase.

14. General Disclaimer

This article is an educational explanation of the research paper “Automatic Classification and Color Changing of Saree Components using Deep Learning Techniques.” It is intended for conceptual understanding, academic discussion, and research learning. Some technical details have been simplified for readability. For exact methodology, implementation details, experimental settings, and complete results, readers should refer to the original paper.

```