Convolutional neural network for dactylological alphabet recognition of Honduran Sign Language (LESHO)

Introduction The critical problem of hearing and speech difficulties among thousands of Hondurans is a latent need, and in search of tools that allow the inclusion of the population with this condition is sought through the introduction of a convolutional neural network (CNN) designed for real-time detection and classification of the dactyl alphabet of the Honduran sign language (LESHO). This study represents an important step forward in promoting accessibility and inclusion of the Honduran deaf community, which faces few technological solutions adapted to their needs. Methods A proprietary dataset comprising more than 8,000 images with various angles and gestures was meticulously constructed, ensuring robust training and evaluation of the model. Spiral research methodology was employed to iteratively refine network performance, with an emphasis on accuracy and real-time deployment capabilities. Results The final model showed exceptional results during the testing, achieving a mean average precision (mAP) of 98.8%, a precision of 97.4%, and a recall of 97.7%. These metrics underscore the reliability of the CNN in recognizing both static and dynamic gestures with minimal errors. Conclusion The model’s capacity to generalize indicates its potential for further applications, such as full sign language interpretation and expanded vocabulary training