Fine-grained recognition of cultural artifacts remains challenging because of the scarcity of annotated data, subtle intra-class differences, and heterogeneous imaging conditions. This study addresses these issues through a domain-specific deep learning pipeline, demonstrated on Indonesian keris classification across three tasks: pamor (27 classes), dhapur (42), and tangguh (5). The pipeline integrates background homogenization, orientation normalization, and YOLOv8-based blade cropping with mask-aware augmentation restricted to the blade regions. For classification, we propose KerisRDNet, which extends InceptionResNetV2 with Inception-Residual-Dilated (IRD) blocks and squeeze-and-excitation to model the elongated geometries and subtle forging motifs. Experiments show that baseline networks collapse under fine-grained settings, with macro-F1 near zero, whereas the proposed approach achieves 0.268 ( pamor ), 0.276 ( dhapur ), and 0.635 ( tangguh ) with Top-3 accuracy above 0.5 and AUC up to 0.853. Across three stratified resamplings, paired non-parametric tests (Wilcoxon signed-rank) indicated directionally consistent improvements; given the small number of repetitions ( n = 3 ), these results are interpreted conservatively. These results demonstrate the feasibility of practically viable keris recognition as a decision-support tool for cultural heritage curation, while also offering a transferable workflow for low-data fine-grained recognition tasks. • Domain-specific deep learning pipeline for fine-grained artifact recognition. • Mask-aware augmentation preserves blade integrity and balances imbalanced data. • KerisRDNet extends InceptionResNetV2 with Inception–Residual–Dilated blocks. • Achieves macro-F1 up to 0.635 and AUC 0.853 across three keris classification tasks. • Provides reproducible, transferable framework for intelligent cultural heritage systems.
Hastuti et al. (Thu,) studied this question.