<![CDATA[This study aims to examine the extent to which segmentation can improve the classification accuracy of Batak Toba Ulos motifs and to compare the effectiveness of three Convolutional Neural Network (CNN) architectures, namely VGG16, Inception-V3, and MobileNetV3, in classifying the segmentation results. The study is motivated by the high similarity of patterns, colors, and textures among ulos motifs, as well as visual noise from background, lighting, and shadows that reduce classification accuracy. The method consists of two main stages, segmentation and classification. Segmentation begins with manual polygon annotation using VGG Image Annotator (VIA), converted into COCO format as ground truth to train a Mask R-CNN model, which then separates the motif area from the background, producing a Region of Interest (ROI) as input for classification. The dataset consists of 700 images of seven types of Batak Toba ulos obtained through direct image acquisition using a smartphone camera. Evaluation used the mean Average Precision (mAP) metric for segmentation, and accuracy, precision, recall, and F1-Score for classification. The results show that Mask R-CNN segmentation is effective, achieving a Mean IoU of 0.9062, a Bounding Box AP of 0.9228, and a Segmentation AP of 0.8808. In classification, all three CNN architectures achieved accuracy above 98%, with VGG16 and MobileNetV3 reaching the highest accuracy of 99.71%, while Inception-V3 achieved 98.57%. In terms of computational efficiency, MobileNetV3 is the most recommended architecture, as it matches VGG16's accuracy with far fewer parameters and a shorter training time.]]>
Copyrights © 2026
