Dataset Transformation System for Sign Language Recognition Based on Image Classification Network

Among the various fields where deep learning is used, there are challenges to be solved in motion recognition. One is that it is difficult to manage because of the vast amount of data. Another is that it takes a long time to learn due to the complex network and the large amount of data. To solve the problems, we propose a dataset transformation system. Sign language recognition was implemented to evaluate the performance of this system. The system consists of three steps: pose estimation, normalization, and spatial-temporal map (STmap) generation. STmap is a method of simultaneously expressing temporal data and spatial data in one image. In addition, the accuracy of the model was improved, and the error sensitivity was lowered through the data augmentation process. Through the proposed method, it was possible to reduce the dataset from 94.39 GB to 954 MB. It corresponds to approximately 1% of the original. When the dataset created through the proposed method is trained on the image classification model, the sign language recognition accuracy is 84.5%.