Ornament image retrieval using few-shot learning.

In this paper, we introduce OrnamentFIR, a novel ornament dataset related to the fashion industry. In recent years, the retrieval of clothing and footwear articles has received significant interest from researchers. However, because of the design intricacy and lack of a suitable dataset, intricate fashion products, like jewelry, have not gotten much attention. We have assembled the OrnamentFIR dataset to address this issue. By revisiting the publically accessible datasets, namely RingFIR and NecklaceFIR, we create a novel dataset called OrnamentFIR. The dataset includes over similar to 4.4 K high-quality images of bangles, over similar to 4.8 K high-definition images of necklaces, and more than similar to 2.6 K high-quality images of earrings. The dataset is divided into three named classes: ring, necklace, and bangle, with each class having 46, 49, and 56 labeled categories, respectively. Due to the limited amount of data, we employed matching networks, a neural network that uses recent advances in attention and memory to enable rapid learning, to extract the desired image from the dataset. Using the matching networks for one-shot learning technique, we achieve 68% accuracy for RGB photographs, 62% accuracy for segmented images, and 50% accuracy for RGB+Segmented images. For the benefit of researchers, the ornament dataset has been made public. Public access to the dataset and code is provided at https://github.com/iammaidul/OrnamentFIR.