Bhutanese Handwoven Textile Pattern Recognition and Classification Using Artificial Neural Networks

Abstract

Bhutanese textiles symbolize the country's unity and autonomy through cultural aesthetics and distinctive features of traditional wear. The weavers and the experts can distinguish and recognize the differences in patterns instantly. However, most Bhutanese individuals, particularly the youths and the foreigners find it difficult to recognize the differences in it. Identification of such unique textiles is often learnt through practice, making it a particularly rigorous and costly form of learning, resulting in discontent among people who perform it. Bhutan also has a scarcity of digital archives and paper materials for textile identification and future references. This study aims to develop machine learning models that are capable of recognizing and classifying the Bhutanese textile types and their’ patterns using Bhutanese textile datasets.

In this study, the first Bhutanese textile datasets were developed in this work using various augmentation approaches. Bhutanese handwoven textile pattern (BHTP) and Bhutanese textile types (BTT) were generated as two separate datasets. The datasets were used to test various textile recognitions models.

A PatternNet model was proposed for recognizing and classifying Bhutanese handwoven textile patterns into 10 classes. Using the CNN (Convolutional Neural Network) concept, a six-layered PatternNet model was developed and trained on the BTP dataset. PatternNet model outperformed other models such as VGG16, AlexNet, ResNet-34, ResNet-50, SVM, and KNN with an accuracy of 99.75% and 99.25% for training and validation respectively. Similarly, VGG16 architecture outperformed KNN, SVM and AlexNet in recognition and classification of 7 types of Bhutanese textiles with the accuracy of 99% and 98.33% for training and validation respectively. PatternNet model was deployed on the flask web application and it was tested with 550 input images. The testing accuracy of the PatternNet model deployed on web application was 98.54%.

This is the first study to emphasize the feasibility of using machine learning to classify the types and patterns of Bhutanese textiles. Machine learning techniques have been shown to be effective in the development of applications of Computer Vision to make Bhutanese textile learning openly accessible. This study can be expanded to detect the defects present in textiles and generate new textile patterns or motifs.

-