DETECTION OF MOLLO WOVEN FABRIC MOTIFS FROM EAST NUSA TENGGARA USING A SIMILARITY AND OBJECT DETECTION APPROACH

• 主要作者: Irawan Firdaus, Ferdy
• 格式: Final Project
• 語言: Indonesia
• 在線閱讀: https://digilib.itb.ac.id/gdl/view/78146
• 機構: Institut Teknologi Bandung

According to the Big Indonesian Dictionary, woven cloth is a work of art that uses thread in the manufacturing process. Woven fabric is one of Indonesia's cultural heritages originating from various regions, one of which is Mollo. Mollo is one of the regions in South Central Timor. Mollo woven fabric can be differentiated based on two factors, namely the manufacturing technique and the motif of the woven fabric itself. Today machine learning has penetrated various areas of human life, one of which is culture and art. This can be used to identify woven fabric motifs. The problem is that datasets are still difficult to obtain, apart from that, Mollo woven cloth motifs also increase over time due to religious and cultural processes, so it is possible that in one woven cloth there are several woven motifs because of this. Related research uses an object detection approach with the Faster R-CNN model. In this research, a model was built using a similarity and object detection approach. The most suitable approach for this research will be selected and the best model will be selected based on its performance metrics. The similarity model built is Siamese network and the object detection models built are Faster R-CNN and YOLO. The Siamese network similarity model that was built gave good results for relatively small datasets, the contrastive loss result was 0.003368. However, the drawback is that it cannot handle the problem of more than one motif object in one woven fabric, the angle of taking the image greatly influences the prediction results, and the output is still the distance between the similarities between the two motifs. The object detection model was built to overcome problems that cannot be overcome by the Siamese network. The object detection models built include Faster R-CNN X101-FPN and YOLOv8. Hyperparameter tuning is carried out to obtain the most optimal parameter configuration. The performance metrics used are mAP50, precision, recall, and inference time. The mAP50 performance metric is a popular metric to express the accuracy of an object detection model. Precision measures the accuracy of the model in classifying samples as positive class. Recall measures the model's ability to detect Positive samples (model sensitivity). Inference time is the time needed for the model to process data to make predictions. The YOLOv8 model was chosen because it produces the highest mAP50, precision and recall, namely 96.1%, 91.397% and 92.026% respectively. The fastest YOLOv8 inference time is 2.8ms.