Braille Recognition of Geez Numbers Using an Optimal Deep Learning Algorithm
Bekalu Tadele Abeje, Ayodeji Olalekan Salau, Gebeyehu Belay, Gunjan Chhabra, Keshav Kaushik, Sepiribo Lucky Braide · 2024
Most of the time, people use their voices for communication and their vision to interact with the external environment. There are people who do not have the opportunity to use their vision to interact with the external environment due to different factors of disability. One of those community groups is the blind people. Blind people can communicate with other people with the help of their sounds, but they cannot communicate and exchange messages by means of writing. To overcome this challenge, the braille writing system is the prior one. Braille is the only way that they can communicate with their peers as well as other community members in the form of writing. People around the world use various languages for communication, and Geez is one of the ancient and popular languages of Ethiopians, particularly in the Ethiopian orthodox tewahedo church, where it is used for many aspects of spiritual worship. There have been numerous studies on Geez and Amharic alphabet braille recognition, but none of that research was conducted on Geez number braille recognition. This study focuses on recognizing geez braille signs into geez number digital written format by using braille image signs as input and geez numbers as the desired output. Serious steps including image acquisition, preprocessing, image segmentation, data augmentation, feature extraction, and recognition were applied. We have collected 1900 JPG images from hardcopy documents using iPhone 6 s camera and scanner, which belongs to the 19 class. Our model is trained with 100 epochs by partitioning the dataset into 0.8% of data as training and 0.2% as testing data. We have used end-to-end convolutional neural networks to recognize Geez numbers from ፩-፲ and base numbers up to ፻. The proposed model achieves 96.88% training accuracy, 95.26% validation accuracy, 0.461% training loss, and 0.2844 validation loss recognition performance. The proposed model was optimal in the case of overfitting and underfitting phases as shown in the result achieved.