Pengenalan gaya seni pada lukisan dalam konteks computer vision memiliki tantangan karena perbedaan visual yang kompleks, kesamaan ciri antar-gaya, dan karakter abstrak beberapa gaya seni, sehingga klasifikasi menjadi sulit. Eksperimen awal dengan enam kelas (Ekspresionisme, Impresionisme, Romantisisme, Realisme, Simbolisme, dan Cubisme) menunjukkan bahwa fine-tuning DenseNet pada 10% lapisan awal, tengah, dan akhir masing-masing menghasilkan akurasi validasi 55,83%, 55,50%, dan 63,67%, dengan precision, recall, dan F1-score tertinggi dicapai pada lapisan akhir (66,18%, 63,67%, dan 63,97%). Beberapa kelas seperti Realisme dan Simbolisme memiliki prediksi rendah, sementara Cubisme berbeda karakter, sehingga dilakukan pengurangan menjadi tiga kelas (Ekspresionisme, Impresionisme, Romantisisme) untuk meningkatkan performa. Pada tiga kelas, DenseNet memperoleh train accuracy 83,45%, validation accuracy 79,67%, train loss 0,4533, validation loss 0,5593, dan generalization gap 0,11, menunjukkan kemampuan generalisasi terbaik dibanding Xception dan ResNet50. Hasil penelitian ini menegaskan bahwa fine-tuning 10% lapisan terakhir pada DenseNet mampu meningkatkan kemampuan model dalam membedakan gaya seni dengan tepat.
Art style recognition in paintings poses significant challenges in computer vision due to complex visual variations, similarities between styles, and the abstract characteristics of some art movements, making accurate classification difficult. Initial experiments using six classes (Expressionism, Impressionism, Romanticism, Realism, Symbolism, and Cubism) showed that DenseNet fine-tuned on the first, middle, and last 10% of layers achieved validation accuracies of 55.83%, 55.50%, and 63.67%, respectively, with the highest precision, recall, and F1-score observed on the last layers (66.18%, 63.67%, and 63.97%). Some classes, such as Realism and Symbolism, consistently had low prediction scores, while Cubism differed in characteristics, leading to a reduction to three classes (Expressionism, Impressionism, Romanticism) to improve performance. For these three classes, DenseNet achieved train accuracy 83.45%, validation accuracy 79.67%, train loss 0.4533, validation loss 0.5593, and a generalization gap of 0.11, indicating the best generalization compared to Xception and ResNet50. These results demonstrate that fine-tuning the last 10% of DenseNet layers effectively enhances the model’s ability to distinguish art styles accurately.
