Thai Silk Patterns Classification with Deep Neural Networks

Abstract

The art of silk weaving has been transferred through generations as part of folk wisdom. Every locality has its distinct silk pattern design. Expertise and familiarity with silk are necessary for the classification of silk patterns. Therefore, only a few experts can recognize the silk's pattern. This study aims to implement a system for classifying silk patterns using image processing technology to help identify silk patterns from images. This research collected silk pattern data from the Chonnabot district, Khon Kaen Province. We selected 15 silk patterns and collected a total of 2,156 images. We examined two convolutional neural networks (CNNs), which differed in feature extraction and regularization via the dropout technique. The experimental results showed that CNN model 1 achieved an F1-score of 0.62. The CNN model 2, in which feature extraction using the pre-trained model was added to the CNN model 2, achieved an F1-score of 0.92, which can assist in resolving the confusion in silk pattern classification.

Keywords

• neural network
• deep neural network
• convolutional neural network
• silk pattern
• classification