Penerapan Domain Adaptation untuk Segmentasi Semantik pada Infrastruktur Jalan di Surabaya

Segmentasi semantik merupakan salah satu teknik utama dalam pengolahan citra digital yang bertujuan untuk mengklasifikasikan setiap piksel ke dalam kategori objek tertentu dan banyak diterapkan pada sistem kendaraan otonom serta pemantauan infrastruktur jalan. Namun, pada penerapan di dunia nyata, kinerja model segmentasi sering mengalami penurunan akibat adanya perbedaan karakteristik visual antara data sintetis yang digunakan pada tahap pelatihan dan kondisi aktual di lapangan. Penelitian ini menerapkan pendekatan domain adaptation pada model segmentasi semantik SegFormer-R101 untuk meningkatkan performa segmentasi citra infrastruktur jalan di Kota Surabaya. Data dunia nyata diperoleh melalui proses web scraping dari Google Street View dan dianotasi secara manual agar merepresentasikan karakteristik jalan perkotaan di Indonesia secara akurat. Proses adaptasi domain dilakukan melalui transfer learning dan fine-tuning, serta penyelarasan distribusi fitur antara domain sumber dan domain target menggunakan metode Optimal Transport. Evaluasi kinerja model dilakukan menggunakan metrik Mean Intersection over Union (mIoU) dan akurasi piksel. Hasil eksperimen menunjukkan bahwa penerapan domain adaptation mampu meningkatkan performa segmentasi citra jalan di Surabaya dibandingkan dengan model tanpa adaptasi domain. Meskipun demikian, model masih mengalami keterbatasan dalam mengenali objek berukuran kecil serta kelas dengan distribusi data yang tidak seimbang.

=====================================================================================

Semantic segmentation is one of the main techniques in digital image processing that aims to classify each pixel into a specific category of objects and is widely applied to autonomous vehicle systems as well as road infrastructure monitoring. However, in real-world applications, the performance of segmentation models often decreases due to differences in visual characteristics between the synthetic data used in the training stage and actual conditions in the field. This research applied a domain adaptation approach to the SegFormer-R101 semantic segmentation model to improve the performance of road infrastructure image segmentation in Surabaya. Real-world data was obtained through a web scraping process from Google Street View and manually annotated to accurately represent the characteristics of urban roads in Indonesia. The domain adaptation process was carried out through transfer learning and fine-tuning, as well as the alignment of feature distribution between the source domain and the target domain using the Optimal Transport method. Model performance evaluation was conducted using Mean Intersection over Union (mIoU) and pixel accuracy metrics. The results of the experiment showed that the implementation of domain adaptation was able to improve the performance of road image segmentation in Surabaya compared to the model without domain adaptation. However, the model still had limitations in recognizing small objects and classes with unbalanced data distribution.