Share
Export Citation
Classification of Toraja Woven Fabric Motifs Using Faster R-CNN with Kolmogorov-Arnold Network
Minggu Y.
Proceedings 2026 International Conference on Informationtechnology and Computing Icitcom 2026
Abstract
This study, connecting the Kolmogorov-Arnold Network (KAN) into the Faster R-CNN architecture is proposed to improve the detection and classification ability of Toraja weaving motifs. The fully connected layer on the Faster R-CNN is replaced by KAN which uses the spline-base activation function so that more complex non-linear relationships can be captured and more complex visual patterns can be processed better than the model. The dataset used consists of 2,837 descriptions of Toraja weaving motifs that have gone through preprocessing and augmentation stages. Although the training loss of Faster R-CNN-KAN is higher at 0.0712 compared to 0.0421 in Faster R-CNN, which is the result of an experiment that indicates an increase in mAP of 1.5% from 0.731 to 0.746. In addition, the AP50 value increased from 0.960 to 0.971 and AP75 from 0.899 to 0.909. These results indicate that despite having higher training losses, the performance of Faster R-CNN-KAN is more effective in detecting and classifying objects with better accuracy on new data. Therefore, KAN can be considered as a promising alternative in improving the detection performance of traditional textile motifs.