East Nusa Tenggara Weaving Image Retrieval Using Convolutional Neural Network

Abstract

The popularity of East Nusa Tenggara (ENT) province is attributed to a variety of traditional woven fabrics with local cultural attributes. Each tribe in the province has its design and colors that differentiate the fabrics leading to diverse decorative motifs. Due to different varieties, it is challenging for users to know both the type of motif and its origins. In this research, several Convolutional neural network (CNN) architecture benchmarks were carried out for ENT weaving images retrieval. The image retrieval method was chosen for the study since it has feature extraction and similarity measurement, which make searching and selection relatively easier. Furthermore, the CNN method is often used for feature extraction due to its ability to recognize objects while hashing and hamming distance algorithms help reduce the computation time for similarity testing. This study was conducted by comparing several pre-trained CNN architectures, including VGG16, ResNet101, InceptionV3, and Discrete Wavelet Transform. The results showed that the highest accuracy is ResNet101 architecture with 100%, 88.50%, and 55% at top=1, top=5, and top=10, respectively. The pre-trained CNN model and Discrete Wavelet Transform combination provided better results in case the feature dimensions were above 16-bit. The feature dimensions are generally based on the best 6-bit hashing code, though they are computationally time-consuming.