<![CDATA[This study re-examines Batik Nitik classification using a leakage-safe provenance-aware evaluation protocol to determine which handcrafted descriptors make a substantive contribution to performance and whether saturated results persist after provenance-based partitioning. Batik Nitik 960 was represented using Histogram of Oriented Gradients (HOG), Local Binary Patterns (LBP), Gray-Level Co-occurrence Matrix (GLCM) descriptors, and grayscale intensity moments. Descriptor ablation, classifier benchmarking, cosine-similarity baselines, four-setting leave-one-provenance-group-out sensitivity analysis, and a supplementary image-level split comparison were evaluated using in-pipeline preprocessing. All HOG-containing feature sets achieved 0.9833 cross-validation accuracy and 1.0000 hold-out accuracy. On fused features, SVM, KNN (Euclidean), and KNN (cosine) achieved 1.0000 hold-out accuracy, while Random Forest reached 0.9958. Raw-pixel, HOG-only, and fused-feature cosine baselines also reached 1.0000 hold-out accuracy. A supplementary image-level HOG-SVM split also produced 1.0000 accuracy. This study contributes a provenance-aware benchmark diagnosis for Batik Nitik classification by identifying HOG as the strongest standalone handcrafted descriptor and by cautioning against deployment-ready interpretation of saturated closed-set accuracy.]]>
Copyrights © 2026
