| Abstrak/Abstract |
This study aims to develop the TenunIkatNet
dataset as one of the benchmarks for image recognition and
retrieval problems, especially in Indonesian traditional fabrics.
These are due to the present absence of the dataset for
traditional fibers, especially the tenun ikat (woven) fabrics of
East Nusa Tenggara (ENT). The TenunIkatNet dataset contains
120 types of fabric each being captured 40 times, leading to the
collection of 4,800 original images at traditional shops and
craftsmen. These shooting variations are often internally and
externally carried out within the mini-studio box and another
fabric background, before being hanged, and worn on the body.
In this condition, the utilized Nikon D5600 camera produced 24
megapixels of RGB (red, green, blue) images, with the pre-
processing and visual procedural application obtaining a
256x256 display according to the input feature extraction
method. Furthermore, the data augmentation processes such as
flipping, random rotating, zooming, and shear range, increased
the number of images, with each type of tenun ikat fabric being
randomly augmented by two images. This led to the production
of 9,600 images based on the number of training datasets.
Despite this, the testing and validation data were 960 images
each. The TenunIkatNet dataset is expected to develop image
recognition, classification, and retrieval algorithms, while also
being useful for the preservation of local culture, as a knowledge
base in the fields of education, crafts, and trade. In this report,
two pre-trained Convolution Neural Network (CNN) methods
with a transfer learning procedure, namely ResNet101 and
DenseNet201. The results showed that the retrieval accuracy of
two pre-trained models is 100% in the top-1. While the retrieval
accuracy of DenseNet201 outperformed the ResNet101 method
in the top-5 at 44.75% and 44.38%, respectively. Subsequently,
the method had higher retrieval accuracy than ResNet101,
although the query time was slower |
