Electronic Theses and Dissertation

Penyebaran iklan judi online berbahasa indonesia menimbulkan risiko finansial dan sosial, sementara pemblokiran berbasis domain menghadapi keterbatasan karena domain baru dapat dibuat secara cepat. penelitian ini bertujuan membandingkan dua pendekatan klasifikasi iklan judi online, yaitu pendekatan visual menggunakan convnext-large dan pendekatan tekstual menggunakan pipeline paddleocr dan xlm-roberta. dataset penelitian terdiri atas 3.000 citra yang dikumpulkan secara mandiri melalui proses crawling dan scraping, dengan komposisi seimbang antara 1.500 citra iklan judi online dan 1.500 citra iklan non-judi. data dibagi secara stratified menjadi data latih, data validasi, dan data uji. convnext-large digunakan untuk mempelajari fitur visual seperti warna, bentuk, objek, dan tata letak iklan, sedangkan pipeline paddleocr dan xlm-roberta digunakan untuk mengekstraksi serta mengklasifikasikan informasi tekstual pada citra. model terbaik dipilih berdasarkan kinerja pada data validasi, kemudian dievaluasi pada data uji yang terdiri atas 450 citra menggunakan metrik accuracy, precision, recall, dan f1-score. hasil pengujian menunjukkan bahwa pipeline paddleocr dan xlm-roberta mencapai akurasi 98,44% dengan f1-score makro 0,9844, sedangkan convnext-large mencapai akurasi 97,33% dengan f1-score makro 0,9733. dari sisi waktu inferensi, convnext-large memproses setiap citra dalam rata-rata 26,94 ms, sedangkan pipeline paddleocr dan xlm-roberta memerlukan rata-rata 1.119,56 ms per citra berdasarkan pengukuran pada data uji. analisis kesalahan memperlihatkan bahwa convnext-large lebih cenderung keliru ketika tampilan visual antarkelas sangat mirip, sedangkan pipeline paddleocr dan xlm-roberta lebih cenderung keliru ketika informasi teks yang berhasil diekstrak dari citra terbatas. convnext-large memiliki waktu inferensi sekitar 41,6 kali lebih cepat daripada pipeline paddleocr dan xlm-roberta, sedangkan pipeline paddleocr dan xlm-roberta memiliki f1-score 1,11 poin persentase lebih tinggi dibandingkan convnext-large.

The spread of Indonesian-language online gambling advertisements poses financial and social risks, while domain-based blocking faces limitations because new domains can be created rapidly. This study compares two approaches for classifying online gambling advertisements: a visual approach using ConvNeXt-Large and a textual approach using a PaddleOCR and XLM-RoBERTa pipeline. The research dataset consists of 3,000 images collected independently through crawling and scraping, with a balanced composition of 1,500 online gambling advertisement images and 1,500 non-gambling advertisement images. The data were split in a stratified manner into training, validation, and test data. ConvNeXt-Large was used to learn visual features such as color, shape, objects, and advertisement layout, while the PaddleOCR and XLM-RoBERTa pipeline was used to extract and classify textual information from images. The best model configuration was selected based on validation performance and then evaluated on a held-out test set consisting of 450 images using accuracy, precision, recall, and F1-score metrics. The test results show that the PaddleOCR and XLM-RoBERTa pipeline achieved 98.44% accuracy with a macro F1-score of 0.9844, while ConvNeXt-Large achieved 97.33% accuracy with a macro F1-score of 0.9733. In terms of inference time, ConvNeXt-Large processes each image in an average of 26.94 ms, while the PaddleOCR and XLM-RoBERTa pipeline requires an average of 1,119.56 ms per image based on measurements on the test data. Failure analysis reveals that ConvNeXt-Large is more likely to err when the visual appearance across classes is very similar, whereas the PaddleOCR and XLM-RoBERTa pipeline is more likely to err when the text information successfully extracted from the image is limited. ConvNeXt-Large has an inference time approximately 41.6 times faster than the PaddleOCR and XLM-RoBERTa pipeline, while the PaddleOCR and XLM-RoBERTa pipeline has a macro F1-score 1.11 percentage points higher than ConvNeXt-Large.