Abstract:

Sign language is a way for hearing-impaired people to communicate among themselves and with people without hearing impairment. Communication with the sign language is difficult because few people know this language and the language does not have universal patterns. Sign language interpretation is the translation of visible signs into speech or writing. The sign language interpretation process has reached a practical solution with the help of computer vision technology. One of the models widely used for computer vision technology that mimics the work of the human eye in a computer environment is deep learning. Convolutional neural networks (CNN), which are included in deep learning technology, give successful results in sign language recognition as well as other image recognition applications. In this study, the dataset containing 2062 images consisting of Turkish sign language digits was classified with the developed CNN model. One of the important parameters used to minimize network error of the CNN model during the training is the learning rate. The learning rate is a coefficient used to update other parameters in the network depending on the network error. The optimization of the learning rate is important to achieve rapid progress without getting stuck in local minimums while reducing network error. There are several optimization techniques used for this purpose. In this study, the success of four different training and test processes performed with SGD, RMSprop, Adam and Adamax optimizers were compared. Adam optimizer, which is widely used today with its high performance, was found to be the most successful technique in this study with 98.42% training and 98.55% test accuracy.

Keywords: