Convolutional Neural Network Models for HandWoven Fabric Motif Recognition

Motifs of hand-woven fabrics are normally designed uniquely within each village or area. Thus, they represent story, culture, or history of the related areas. This research is intended to automate the motif recognition. To complete this, the authors surveyed local fabrics in Loei province, Thailand, which is called phasin. Phasins are commonly woven and worn by women. In this process, the authors collected images of phasins based on two types of phasins found in the area called mee-kun and mee-ruad. The collected images were segmented. The segmentation revealed 25 motifs, which were used for experiment. In the experiment, 10-fold cross-validation research design was employed to train and test 4,500 images in the dataset based on two convolutional neural network models: Inception-v3 and MobileNets. The results reveal that MobileNets outperforms Inception-v3 with the accuracy rate of 98.223% and 93.208% respectively.