Ixchel Features Rich, Vibrant Images Inspired by Maya Textiles

Why Maya Textiles? For centuries, Maya women in Guatemala, El Salvador, Honduras, Mexico, and Belize have been manually weaving clothing and other textile arts that encapsulate their beliefs, culture, rituals, and traditions.

Worldwide, however, the knowledge of Maya culture and its rich colors and designs is in danger. High-fashion designers have been known to document this ancient weaving process to later on incorporate the knowledge into creating expensive products giving little to no credit to the artisans.

Traditional indigenous garments worn by Maya people are more than just clothing, they unite and express the identity of ancestral communities.

How to learn more? Initiatives such as Museo Ixchel del Traje Indigena aim to educate the public on the evolution of indigenous dress in diverse communities and the techniques used in its elaboration. Other groups such as AsociaciĆ³n Femenina para el Desarrollo de SacatepĆ©quez feature indigenous women organized to reclaim identity, heritage, and rights, and formally demand intellectual property rights to protect their ancestral designs and weavings. You can also learn and help by visiting indigenous communities all over Central America, and buy textiles directly from them.

TL;DR of our approach

As training set we use free pictures of Maya textiles available in Google Images, we then apply Data Augmentation [1] to transform and increase the training set. We use a Generative Adversarial Network or GAN to implement our approach.

Models. After the Generator and Discriminator models are trained, we save them independently.

From random noise to images. We then use the Generator to produce images from random noise.

Super-Resolution. Finally, we use a tailor-made Non-Adversarial Super-Resolution model based on a dynamic U-Net architecture [2]. The approach is inspired by fast.ai’s ‘Decrappification’ technique [3].

[1] Data Augmentation: https://docs.fast.ai/vision.transform.html#Data-augmentation

[2] U-Net: Convolutional Networks for Biomedical Image Segmentation. Olaf Ronneberger, Philipp Fischer, Thomas Brox. Medical Image Computing and Computer-Assisted Intervention (MICCAI), Springer, LNCS, Vol.9351: 234–241, 2015, available at arXiv:1505.04597

[3] Decrappification, DeOldification, and Super Resolution. Jason Antic (Deoldify), Jeremy Howard (fast.ai), and Uri Manor (Salk Institute). 2019. https://www.fast.ai/2019/05/03/decrappify/