<![CDATA[This study aims to detect Braille letter patterns using the InceptionV3 architecture combined with the application of median filter and image segmentation. The dataset consists of 4,160 Braille images, with an average of 160 images for each letter from A to Z. The data is divided into 3,900 images for training, which are then split into 3,120 images for training and 780 images for validation, and 260 images are used for testing. Each image is resized to 299x299 pixels before being fed into the model. This study uses 100 epochs and applies early stopping to avoid overfitting. Two learning rate values are tested, namely 0.001 and 0.0001. The results show that the application of a median filter and segmentation significantly improves model performance, producing better accuracy, precision, recall, and F1 values compared to models without these techniques. At a learning rate of 0.001, the model achieves 99.65% accuracy, 99.62% precision, and 99.61% recall. On the other hand, without a median filter and segmentation at a learning rate of 0.0001, although accuracy and precision decreased, the values still reached 99.65% and 99.62%.]]>
Copyrights © 2025
