A Novel Optimization Strategy for CNN Models in Palembang Songket Motif Recognition

Palembang Songket is an essential part of Indonesian cultural heritage, and its introduction and preservation present challenges, particularly in recognizing various motifs. This research introduces a novel strategy to optimize the performance of Convolutional Neural Networks (CNNs) by presenting a hierarchical integration of Ghost Module operations and Max Pooling, referred as Ghost Feature Maps. While the Ghost Module is effective in reducing parameters and enhancing feature extraction, it has limitations in filtering irrelevant information. To address this shortcoming, we propose a hierarchy in which Max Pooling works in conjunction with the Ghost Module, strengthening its performance by not only extracting dominant features but also eliminating excess, non-essential information. This hierarchical design enables more efficient feature extraction, thus enhancing the model's recognition accuracy. By combining Ghost Modules and Max Pooling in a structured manner, this approach advances established methodologies and offers a new perspective on feature representation within CNN architectures. Utilizing a dataset of 10 augmented classes of Palembang Songket motifs totaling 1000 images, we conducted experiments using varying ratios of Ghost Feature Maps. The results indicate that a ratio of 2 achieves an impressive accuracy of 0.98 with minimal parameter reduction. Additionally, a ratio of 3 results in a 34% decrease in parameters while maintaining a competitive accuracy of 0.95. Ratios of 4 and 5 continue to demonstrate robust performance, achieving accuracy levels of 0.93 while delivering over 60% reductions in model size and parameters. This research not only contributes to the optimization of CNN architectures but also supports the preservation of cultural heritage by improving the recognition capabilities of Palembang Songket motifs.

@article{Yohannes2025,
  title     = {A Novel Optimization Strategy for CNN Models in Palembang Songket Motif Recognition},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {16},
  number    = {1},
  year      = {2025},
  publisher = {The Science and Information Organization},
  author    = {Yohannes and Muhammad Ezar Al Rivan and Siska Devella and Tinaliah},
  doi       = {10.14569/IJACSA.2025.0160180},
  url       = {https://doi.org/10.14569/IJACSA.2025.0160180}
}