Evidence record · UDINUS · SDGs
- Unit
- Rektorat UDINUS
- Period
- 2026
- Goals
- Record
- 7e173b40-0953-4dae-9bde-b937fb0b481e
- Published
- 2026-10-30
Universitas Dian Nuswantoro (UDINUS), through researchers Hastuti, Khafiizh, Erwin Yudi Hidayat, Abu Salam, and Usman Sudibyo, conducted a study on fine-grained recognition of cultural artifacts. The activity addressed the scarcity of annotated data, subtle intra-class differences, and heterogeneous imaging conditions. It demonstrated a domain-specific deep learning pipeline for Indonesian keris classification across three tasks: pamor, dhapur, and tangguh. The study therefore focused on a computational approach to recognizing cultural heritage blades and supporting cultural heritage curation.
The pipeline integrated background homogenization, orientation normalization, and YOLOv8-based blade cropping. It also used mask-aware augmentation restricted to blade regions. For classification, UDINUS proposed KerisRDNet, an extension of InceptionResNetV2 incorporating Inception-Residual-Dilated blocks and squeeze-and-excitation. These components addressed the elongated geometries and subtle forging motifs described in the source. The classification tasks included 27 pamor classes, 42 dhapur classes, and 5 tangguh classes.
Experiments showed that baseline networks collapsed under fine-grained settings, with macro-F1 near zero. KerisRDNet achieved macro-F1 values of 0.268 for pamor, 0.276 for dhapur, and 0.635 for tangguh. Top-3 accuracy was above 0.5, and AUC reached up to 0.853. Across three stratified resamplings, paired non-parametric Wilcoxon signed-rank tests indicated directionally consistent improvements. Because the number of repetitions was small, with n=3, the results were interpreted conservatively. The study reported the feasibility of practically viable keris recognition as a decision-support tool for cultural heritage curation and offered a transferable workflow for low-data fine-grained recognition tasks.