A new sign language recognition data set called ERUSLR A classification model based on global / heuristic search method and GoogLeNet supported CNN 93.93% accuracy rate obtained by genetic algorithm based classification model Sign language is an important communication tool for deaf-and-dumb people. There is a need to make provision for effective communication in the emergency departments of a hospital. Therefore, firstly, the 25 words needed in emergency services were identified, and subsequently used to create a new dataset, named Erciyes University Sign Language Recognition (ERUSLR) dataset. To classify the ERUSLR dataset, a GoogLeNet-based CNN model with global/heuristic search support is proposed as shown in Figure A. The GoogLeNet pre-trained model was used for transfer learning. A GoogLeNet-based CNN was created by replacing the last 3 layers with suitable layers that are stem from the related problem. Grid search (GS), random search (RS), and genetic algorithm (GA) were used as the parameter search method. These search methods were used to optimize the training options of the CNN including the initial learning rate, ℓ2 regularization, and momentum parameters. The experimental results show that the GA supported model displayed a better performance. Figure A. The proposed system Purpose: The first aim of the study is to create a new sign language recognition dataset that can be used in the hospital. The other aim involves the development of a global/heuristic search and GoogLeNet based method. Theory and Methods: In this study, instead of developing a CNN model from scratch, the transfer learning method, an easier and more effective approach, is selected. Consequently, a GoogLeNet based-CNN model was created by transfer learning from the GoogLeNet pre-trained model. The initial experiments with the ERUSLR dataset created in this study was performed using the GoogLeNet-based CNN model with the default parameter values. Search methods are typically used in parameter optimization to save time. In this study, GS, RS, and GA methods are used to optimize the training parameters of the GoogLeNet-based CNN model. Results: In the experimental studies, the GoogLeNet-based CNN model with the default parameter values recorded an accuracy rate of 88.62%. On the other hand, an accuracy rate of 93.63% was obtained with GS supported GoogLeNet-based CNN model while an accuracy rate of 89.81% was achieved with the RS supported GoogLeNet-based CNN. Finally, an accuracy rate of 93.93% was achieved with the GA supported GoogLeNet-based CNN.
Conclusion:Transfer learning operation is an easy and effective way of building a new CNN model. Global/heuristic search-based parameter optimization is an important factor that improves classification performance.