<![CDATA[This study examines a hybrid method for classifying fine-grained Indonesian batik motifs under limited data conditions. The research focuses on two objectives: (1) assessing the contribution of attention mechanisms to the extraction of discriminative visual features, and (2) evaluating the role of Gray-Level Co-occurrence Matrix (GLCM) texture descriptors when combined with deep convolutional representations. The proposed approach employs a ResNet-50 backbone equipped with a Convolutional Block Attention Module (CBAM) and integrates second-order GLCM features through a feature-fusion framework. The dataset consists of authentic batik photographs representing 38 motif categories. Model performance is assessed using accuracy, macro-averaged metrics, Cohen’s Kappa, and ablation experiments supported by statistical tests. The model reaches a test accuracy of 75.90%, with a macro F1-score of 0.7598 and a Cohen’s Kappa value of 0.7456. Training and validation curves show stable behavior after the initial epochs. Per-class evaluation indicates that motifs with distinctive structural elements tend to be classified correctly, whereas motifs with subtle or overlapping patterns exhibit lower accuracy. The ablation study records a 4.79% accuracy increase attributed to CBAM and a 3.51% increase associated with GLCM features; both effects fall within statistically significant confidence intervals. The combination of both components yields an 8.38% improvement over the baseline model. Two-way ANOVA identifies main effects for attention and GLCM, with a small interaction term. These results provide information on how spatial attention and statistical texture features contribute to the classification of fine-grained batik motifs within the examined setting.]]>
Copyrights © 2025
