1

26-Nov-20

ICCIT 2020

The 23^rd International Conference on Computer and Information Technology

​

​

Title of The Paper

Jamdani Motif Generation Using Conditional GAN

​

Authors

​

​

​

​

​

​

Presented by

Humaira Ferdous Shifa

​

​

• MD Tanvir Rouf Shawon
• Raihan Tanvir
• Humaira Ferdous Shifa
• Susmoy Kar
• Mohammad Imrul Jubair

​

2

26-Nov-20

JAMDANI MOTIF

THE MASTERLY CRAFTMANSHIP

Jamdani, the only surviving cotton variant of Dhakai Muslin, declared as the “Intangible Cultural Heritage of Humanity” by UNESCO in 2013.

​

Jamdani Motifs are:

​

• Geometric exposition of flora & fauna of Bangladesh
• Not sketched on the fabric, but woven directly on the loom from the imagination stored in the minds of Artisans
• Formed by mathematical interlacing of wraps and weft

​

​

​

3

26-Nov-20

Naming The Motifs

Beelpata Phool Paar

Indur Paar

Projapoti Phool

Beelpata Paar

Beelpata Paar

Source: Images collected from the book Traditional Jamdani Designs

4

26-Nov-20

Objective | Simulation of The Proposed System

5

26-Nov-20

6

26-Nov-20

The Butterfly Effect Through Ages

In the Past:

• Jamdani weaver played the role of both as an artisan and a weaver, designing from imagination while weaving on the loom.

7

26-Nov-20

Current Scenario

• No Master Artisan alive now
• Weaver depend on catalogs for weaving motifs on the loom
• Earns BDT 12 per hr.
• Works up to 14 hrs. each day

​

Proposed Solution

Resuscitating the industry through a dedicated tool for Jamdani Motif generation can:

• Bridge the gap between weavers & Designers
• Become an easy and faster way for creating catalogs
• Preserve the surviving motifs
• An international exposure for this industry

Downbeat VS Upbeat

8

26-Nov-20

Current Scenario

Proposed Solution

• A dedicate tool will:
• Act as an intelligent artist
• Make designing easier for enthusiast
• Artificially created motifs will:
• Unveil new source of inspiration for artists
• Keep the visual and artistic appeal of the produced motifs intact

Downbeat VS Upbeat

Motifs are:

• No longer confined to textiles only
• Being used nationally and internationally
• Are made into jewelry, home decor, curtains, utensils, etc.
• Created not by Jamdani weavers but by Artists, fashion designers, entrepreneurs, and craft enthusiasts

9

26-Nov-20

Background Study

• StyleGAN

• A Style-Based Generator Architecture for Generative Adversarial Networks
• Automatically learned, unsupervised separation of high-level attributes and it enables intuitive, scale-specific control of the synthesis

​

• CycleGAN

​

• PixelDTGAN

• Cross-domain transfer GANs will be likely the first batch of commercial applications. These GANs transform images from one domain (say real scenery) to another domain

• Hand-loom Design Generation using Deep Neural Networks

• ESRGAN

10

26-Nov-20

Deep Dive Into Pix2Pix GAN�(Image-to-Image Translation with Conditional Adversarial Networks by Phillip Isola and others)

• Generic
• Does not require to define any relationship between the two types of images
• Makes no assumptions about the relationship
• Learns the objective during training, by comparing the defined inputs and outputs during training and inferring the objective

So we opted to go for a Pix2Pix inspired model!

11

26-Nov-20

Data Collection

Internet

Two authentic sources:

• The Book: Traditional Jamdani Designs
• Direct Observation (A day tour in Jamdani Polli)

​

• Building own dataset
• Ensure Data Authenticity

12

26-Nov-20

The Exhibition

Our Team

• Introduction to the history & heritage of Jamdani
• Observation of motifs from different eras
• Dialogue with artists, designers and weavers
• Collected the the book titles “Traditional Jamdani Designs” an archive of authentic Jamdani motifs, which is one of our sources of data collection

Source: Images taken at Bengal Shilpalay at the Jamdani Festival 2019

13

26-Nov-20

Sample Data Collected From The Archive

Old Jamdani Motifs

Probably from 19^th century

Contemporary Jamdani Motifs

Woven Design & Line Drawing Sample Pair

The Book: Traditional Jandani Design

Source: Images taken from the book Traditional Jamdani Designs

14

26-Nov-20

A Day at Jamdani Polli

Tarabo, Shonargaon

Location: Shahina Jamdani Weaving Factory

Photo: Hand looms used for weaving Jamdani Saree

15

26-Nov-20

Weavers on the loom

Data Collection directly from the loom

Our Team

The Process of Weaving & Data Collection

Preparing a Yarn

Location: Shahina Jamdani Weaving Factory

16

26-Nov-20

Samples of Data Collected From Direct Observation

A saree with modern Jamdani design

getting weaved on the loom

A Saree with traditional

Jamdani motifs

Panjabi with Jamdani

Motifs

Source: Photos of the motifs taken at Shahina Jamdani Weaving Factory

17

26-Nov-20

18

26-Nov-20

Jamdani Boarder Design

Extracted Motifs

Source: Image taken from the book Traditional Jamdani Designs

19

26-Nov-20

Pix2Pix Data Format

This type of data consists of input and desired output side by side :

​

(Boundary to Shoes type)

(Boundary to Handbags type)

20

26-Nov-20

Real Image

Binary Image

Processed Data

Multiplication

skeletonize

Skeleton

Opening/Closing/Dilation

/Erosion

21

26-Nov-20

Five versions of the dataset Jamdani Noksha—Skeleton, Reduced Branch, Sketch, Boundary, and Enhanced Resolution.

22

26-Nov-20

Reduced Branch

Sketch

Skeleton

Boundary

Enhanced Resolution

Samples of different variance of Dataset

23

26-Nov-20

24

26-Nov-20

where G tries to minimize this objective against an adversarial D that tries to maximize it. The generator is tasked to not only fool the discriminator but also to be near the ground truth output in an L2 sense. We also explore this option, using L1 distance rather than L2 as L1 encourages less blurring,

The objective of a conditional GAN can be expressed as,

Our final objective is,

25

26-Nov-20

HOW DOES GAN WORKS?

26

26-Nov-20

27

26-Nov-20

Sample output (middle column of each group) for model trained on Jamdani Noksha’s Boundary, compared to ground truth (right column). Left column shows input strokes from user.

Fig: Loss graph for training on Boundary version

Boundary Version

​

L1 indicates absolute pixel to pixel translation. As the number of iteration increases the L1 tends to get lower, i.e.: generator become more competent in producing images that matches the real image distribution.

28

26-Nov-20

Sample output (middle column of each group) for model trained on Jamdani Noksha’s Enhanced Resolution, compared to ground truth (right column). Left column shows input strokes from user.

Fig: Loss graph for training on Enhanced Resolution version

Enhanced Resolution Version

​

L1 indicates absolute pixel to pixel translation. As the number of iteration increases the L1 tends to get lower, i.e.: generator become more competent in producing images that matches the real image distribution.

29

26-Nov-20

Sample output (middle column of each group) for model trained on Jamdani Noksha’s, (Reduced Branch compared to ground truth (right column). Left column shows input strokes from user.

Fig: Loss graph for training on Reduced Branch version

Reduced Branch Version

​

L1 indicates absolute pixel to pixel translation. As the number of iteration increases the L1 tends to get lower, i.e.: generator become more competent in producing images that matches the real image distribution.

30

26-Nov-20

Sample output (middle column of each group) for model trained on Jamdani Skeleton, compared to ground truth (right column). Left column shows input strokes from user.

Loss graph for training on Skeleton version

Skeleton Version

​

L1 indicates absolute pixel to pixel translation. As the number of iteration increases the L1 tends to get lower, i.e.: generator become more competent in producing images that matches the real image distribution.

31

26-Nov-20

Sample output (middle column) for model trained on Jamdani Noksha’s hand-drawn sketch version, compared to ground truth (right column). Left column shows input strokes from user.

Loss graph for training on Sketch version

Sketch Version

​

L1 indicates absolute pixel to pixel translation. As the number of iteration increases the L1 tends to get lower, i.e.: generator become more competent in producing images that matches the real image distribution.

32

26-Nov-20

Output Analysis

​

Outputs of 5 different versions of dataset

33

26-Nov-20

Loss Graph Analysis

​

Loss graph for training on (a) Boundary, (b) Enhanced Resolution, (c) Reduced Branch, (d) Skeleton, and (e) Sketch version

(a) (b) (c) (d) (e)

34

26-Nov-20

Constraints:

• Insufficiency of data
• The Sketch Version of Jamdani Noksha Dataset has only 250 data. As the sketches are drawn by hand which is a time consuming process large number of data couldn’t be produced.

Future Work:

• Make the outputs more realistic and flawless.
• Classify the Jamdani motifs/patterns from other designs or patterns.
• Generate larger designs using the building block motifs.
• Convert different objects into a geometric pattern that resembles the hand-loomed Jamdani designs.

35

26-Nov-20

​

​

ANY

​

​

QUESTIONS?

36

26-Nov-20

THANK YOU!