Analysis of the Best Optimizer Used by Convolutional Neural Network Algorithms in Detecting Masked Faces

Face mask detection is a way to detect whether someone is wearing a face mask or not. The use of face masks is very important in preventing transmission of the Covid-19 virus. Even though the Covid-19 crisis has ended, the use of masks is still recommended for the public when in crowds, especially when in closed rooms. Therefore, in several meeting buildings there are face mask detection devices. These face mask detection tools are often less accurate in detecting masked faces. This inaccuracy can be caused by less than optimal detection algorithms used such as CNN. The Convolutional Neural Networks (CNN) method is a type of deep learning model that is popularly used in facial mask detection tools and various other types of deep learning-based tools. Through the literature study process, previous research found that CNN can use several optimizers to speed up the training process and increase its accuracy. However, it is not known which optimizer has the most optimal capabilities in the CNN algorithm. Therefore, in this research, a comparison of optimizers was carried out in the model training process using a CNN architecture called VGG 16. VGG16 is a deep neural network architecture consisting of 16 layers. The VGG16 architecture has 13 convolution layers, 2 fully connected layers and 1 classifier layer. In this research, there were three optimizers whose quality was tested, namely SGD, RMSProp, and also Adam. The research results show different effects of each optimizer on model performance, where the Adam optimizer provides the highest accuracy of 97.2%, RMSProp provides an accuracy of 96.92%, and SGD provides the lowest accuracy of 66.85%. Then in the data loss variable, Adam gave the lowest loss of 0.0756, RMSProp gave a loss of 0.0883, and SGD with the highest loss of 0.6124. In this research, it was found that the best optimizer used on CNN with the VGG16 architecture was the Adam optimizer with the best accuracy and data loss results among the other 2 optimizers. This can be seen in the in-depth comparative analysis presented in the paper to help select the right optimizer to increase the effectiveness of face mask detection.