Comparative Analysis of Pre-trained Image Classification Models through Fine-Tuning on Indonesian Traditional Batik Motifs

Abstract

Recognizing Indonesian batik motifs automatically is hard because many classes repeat the same geometric, floral, and symbolic elements, so the visual boundaries between them stay subtle. We build a reproducible transfer-learning benchmark on 13 traditional batik motif classes from 1,305 curated images with a fixed 915/260/130 train/validation/test split. Nine pre-trained classifiers, spanning classic, efficient, and modern CNNs, a Vision Transformer, and YOLO classification pipelines, were fine-tuned in a Docker-based GPU environment, and we add Batik-HybridFusion, a validation-guided, confidence-weighted late fusion of ConvNeXt-Tiny, EfficientNetV2S, and DenseNet169. Batik-HybridFusion and EfficientNetV2S both reach 99.23% test accuracy and 0.9923 macro-F1, and bootstrap confidence intervals together with exact McNemar tests show that the gaps among the leading models aren’t statistically significant on this set. Beyond accuracy, we measure single-image inference latency on GPU and CPU and probe robustness under blur, illumination change, occlusion, rotation, perspective change, and JPEG compression, and turn the results into concrete recommendations for deploying a public-facing batik recognition service.