Ghost-Vanilla Feature Maps: A Novel Hybrid Architecture for Efficient Fine-Grained Songket Motif Classification

South Sumatra songket motifs present a challenging fine-grained classification task due to high inter-class similarity and substantial intra-class variability. This study proposes the Ghost-Vanilla Feature Map, a novel hybrid architecture that integrates low-cost ghost-generated features with the lightweight structural stability of VanillaNet to enhance discriminative feature learning while reducing computational burden. The proposed architecture is designed to address the inefficiency of conventional convolution-heavy networks in capturing subtle motif variations. Experimental evaluation on a dataset comprising 20 songket motif classes demonstrates that a ghost ratio 2 achieves the best trade-off, attaining an accuracy of 0.98 with more than 75% parameter reduction. Increasing the ghost ratio to 3 preserves high classification performance with an accuracy of 0.97, while ratios 4 and 5 further reduce model size at the expense of marginal accuracy degradation. Comparative results indicate that the Ghost-Vanilla Feature Map consistently outperforms lightweight CNN baselines, including MobileNetV3-Small, MobileNetV4-Conv-Small, EfficientNetV2-Small, and ShuffleNetV2. The proposed architecture substantially surpasses the Vanilla-only baseline, which achieves an accuracy of only 0.860 despite requiring 30.19 million parameters, highlighting the limitations of conventional convolution-dominant designs in fine-grained textile classification. The hybrid configuration with a ghost ratio 2 delivers superior accuracy while nearly halving the parameter count and significantly reducing computational overhead. Overall, the Ghost-Vanilla Feature Map provides an efficient and highly discriminative solution for fine-grained songket motif classification, achieving strong performance while substantially reducing model complexity through a balanced hybrid representation.