Classification and Augmentation of African Fabrics Using Convolutional Neural Network and Generative Adversarial Network Models

Abstract

Africans have developed sophisticated textile traditions and various fabrics. Their patterns are original artistic explorations of advanced visual paradigms with dazzling brilliance, intricate composition, and a viscerally appealing rippling effect. However, fabric design in Africa is time-consuming, resource-intensive, and effort-intensive. Moreover, there is a scarcity of well-curated local datasets for African fabrics. Therefore, there is a need for an automated approach to generating more African fabric designs. Generative Adversarial Networks (GANs) are generative modelling that involves automatically identifying and learning the regularities or patterns in input data so that the model can be used to produce new examples that could have been reasonably derived from the original dataset. This study focuses on combining generative models and discriminative models to generate numerous images of African fabrics that would be difficult for a fabric artisan to create. A MobileNetV2 classification model which is a Convolutional Neural Network (CNN) was trained on 1121 images across 8 different classes of fabric designs and StyleGAN2-ADA was employed to generate new African fabric designs for various classes. The performance of MobileNetV2 and StyleGAN2-ADA was evaluated using Accuracy, Precision, Recall, F1 score and Fréchet inception distance (FID). An accuracy score of 0.75 was recorded and new African fabric designs for various classes were generated. African fabrics have a rich history and culture, and using StyleGAN2-ADA with a pre-trained model to generate designs for them is a promising field of research that might produce fresh and inventive designs that build on those traditions.