Electronic Theses and Dissertation

Indonesia memiliki kekayaan budaya yang luar biasa, salah satunya adalah batik yang telah diakui UNESCO sebagai warisan budaya tak benda. Namun, penggunaan pakaian batik di kalangan masyarakat mulai berkurang. Penelitian ini bertujuan mengembangkan sistem pendeteksi otomatis pengguna pakaian batik dan non-batik dengan teknologi visi komputer. Penelitian ini mengimplementasikan dua model deep learning: YOLOV8 untuk deteksi objek orang dan MobileNetV2 untuk klasifikasi pakaian batik dan non-batik. Dataset citra yang digunakan berasal dari berbagai sumber seperti Instagram, Google Chrome, Bing Images, Kaggle, dan koleksi pribadi. Setelah gambar dikumpulkan, proses dilanjutkan dengan pelabelan menggunakan bounding box pada objek pakaian melalui Roboflow, kemudian dilakukan pre-processing data seperti cropping, resize, normalisasi, dan pembagian data. Model dikembangkan menggunakan pustaka TensorFlow dan Keras dengan teknik transfer learning dari MobileNetV2. Model YOLOV8 pre-trained dilatih selama 100 epoch dan menghasilkan mAP (mean Average Precision) sebesar 74,2% untuk deteksi objek orang. Sementara itu, model MobileNetV2 yang digunakan untuk klasifikasi batik dan non-batik, dilatih dengan data augmentasi dan teknik full fine-tuning. Hasil terbaik dicapai pada epoch ke-21 dari konfigurasi awal 30 epoch dan batch size 16, dengan akurasi tertinggi sebesar 97%.

Indonesia has an extraordinary cultural wealth, one of which is batik which has been recognized by UNESCO as an intangible cultural heritage. However, the use of batik clothes among the community began to decrease. This research aims to develop an automatic detection system for batik and non-batik clothing users with computer vision technology. This study implements two deep learning models: YOLOV8 for the detection of people objects and MobileNetV2 for the classification of batik and non-batik clothing. The imagery dataset used comes from various sources such as Instagram, Google Chrome, Bing Images, Kaggle, and personal collections. After the images are collected, the process continues with labelling using a bounding box on the clothing object through Roboflow, then data pre-processing is carried out such as cropping, resizing, normalization, and data sharing. The model was developed using the TensorFlow and Keras libraries with the transfer learning techniques of MobileNetV2. The pre-trained YOLOV8 model was trained for 100 epoches and produced an mAP (mean Average Precision) of 74.2% for the detection of people objects. Meanwhile, the MobileNetV2 model, which is used for the classification of batik and non-batik, is trained with augmented data and full fine-tuning techniques. The best results were achieved in the 21st epoch of the initial configuration of 30 epoches and batch size 16, with the highest accuracy of 97%.