Implementasi Transfer Learning EfficientNetB0 untuk Identifikasi Motif Batik Nitik melalui Aplikasi Android

Ricky Fadel Mohammad, Ir. Nazrul Effendy, S.T., M.T., Ph.D., IPM.; Prof. Ir. Hanung Adi Nugroho, Ph.D., IPM, SMIEEE.

2024 | Skripsi | FISIKA TEKNIK

    Warisan kebudayaan seperti batik, mencerminkan jati diri bangsa melalui nilai dan makna filosofisnya yang telah ditempa sekian lama. Memahami warisan kebudayaan memungkinkan masyarakat untuk lebih mendalami jati diri bangsa yang luhur. Namun, degradasi budaya lokal dan pasifnya masyarakat terhadap terjadinya klaim atas budayanya menjadi tantangan pelestarian warisan budaya untuk saat ini. Pengembangan aplikasi prediksi batik Nitik dapat menjadi salah satu upaya pelestarian warisan budaya yang intuitif untuk digunakan oleh berbagai kalangan masyarakat.

    Sampel-sampel dari dataset Batik Nitik 960 di-train menggunakan metode transfer learning dengan model pre-trained EfficientNetB0. Variasi training berupa implementasi augmentasi data, modifikasi fully connected layers, dropout rate, learning rate, batch size, serta impementasi fine-tuning diterapkan untuk mendapatkan model terbaik.

    Didapati model terbaik dengan nilai akurasi 0,9986, precision 0,9987, recall 0,9986, F1-score 0,9986, dengan ukuran model 15,59 MB. Berdasarkan model ini, aplikasi prediksi batik Nitik berbasis Android berhasil dikembangkan. Aplikasi tersebut telah teruji mampu memprediksi perwakilan sampel gambar setiap kelas batik Nitik menggunakan sensor kamera smartphone.

    The cultural heritage such as batik reflects the society's identity through its enduring philosophical values and meanings. Understanding cultural heritage enables society to delve deeper into its noble identity. However, the degradation of local culture and the passive response of the community towards claims over their culture pose challenges to the preservation of cultural heritage today. The development of the Nitik batik prediction application could be an intuitive effort to preserve cultural heritage, accessible for various segments of society.

    Samples from the Batik Nitik 960 dataset was trained using transfer learning method with a pre-trained EfficientNetB0 model. Training variations included data augmentation implementation, modification of fully connected layers, dropout rate, learning rate, batch size, and fine-tuning implementation to obtain the best model.

    The best model achieved an accuracy of 0.9986, precision of 0.9987, recall of 0.9986, F1-score of 0.9986, with a model size of 15.59 MB. Based on this model, an Android-based Nitik batik prediction application was successfully developed. The application has been tested and proven capable of predicting a representative sample of each Nitik batik class using a smartphone camera sensor.

Kata Kunci : Transfer Learning, EfficientNetB0, Multiclass Classification, Batik Nitik, Android, Artificial Intelligence