Improved pattern recognition of various traditional clothes with Convolutional neural network

Abstract

This Study would identify regional fabric patterns throughout Indonesia. Traditional clothes patterns come from the handicrafts of remote communities in the regions of Indonesia. A Convolutional neural network is used for Resnet50 V2, VGG16, VGG19, MobileNetV2, and Inception V3 models to support this research. This research was conducted on five types of experiments with supporting parameters: the number of training images and test images, rotation and scale of each image, image class, and CNN parameters. After we conducted experiments on a collection of images found in various scales and degrees, the average classification accuracy using VGG16 was 79.2% in condition 2, but the classification accuracy was between 84.2%-92.5%. Meanwhile, if using VGG19 the classification accuracy is 79.9%, with condition 1 for the highest accuracy between 92%-96%. If we used mobilenetV2, the average classification accuracy is 83.4% for 87%-96%. If we used inceptionV3, the average classification accuracy is 83.7% for the accuracy between 72%-98%, and the last if we used Resnet50 V2, the average classification accuracy is 79.2% for the accuracy between 71%-94%. Traditional clothes pattern recognition research using Deep CNN has succeeded in recognizing some traditional clothes patterns up to more than 97%. We continue the study to recognize other traditional clothes patterns more optimally.