<![CDATA[Word segmentation is an essential preprocessing step in offline handwritten text recognition because accurate word localization directly affects subsequent recognition performance. This study presents an Indonesian handwritten word segmentation approach using a neural word detector designed to localize individual words in handwritten student assignment documents. The dataset consists of real handwritten student assignments, with word-level ground truth annotations manually created using LabelMe in the form of rectangular bounding boxes. A pretrained neural word detector was trained for five epochs and evaluated using an Intersection over Union (IoU) threshold of 0.50. Predicted word bounding boxes were matched with ground-truth boxes using the Hungarian assignment algorithm, with the matching cost defined from the IoU between corresponding boxes. Evaluation was conducted on 120 paired files containing 9,267 ground-truth word boxes and 9,308 evaluated predicted boxes. The detector achieved a precision of 92.35%, recall of 92.76%, and F1-score of 92.55%, while the mean IoU of matched bounding boxes reached 86.27%. These results indicate that the neural word detector can effectively localize handwritten Indonesian words with high detection consistency and spatial overlap. The study demonstrates the feasibility of neural word detection as a practical approach for word-level segmentation of real Indonesian handwritten documents. The resulting segmentation can provide localized word regions for subsequent handwritten word recognition and document analysis applications.]]>
Copyrights © 2026
