Learning Rate Analysis for Pain Recognition Through Viola-Jones and Deep Learning Methods

Abstract

Deep learning is growing and widely used in various fields of life. One of which is the recognition of pain through facial expressions for patients with communication difficulties. Viola-Jones is a simple algorithm that has real-time detection capabilities with relatively high accuracy and low computational power requirements. The learning rate is a significant number that has an impact on the deep learning result. This study recognized pain using the Viola-Jones and deep learning methods. The dataset used was a thermal image from the Multimodal Intensity Pain (MIntPAIN) database. The steps taken consisted of segmentation, training, and testing. Segmentation was conducted using the Viola-Jones method to get the significant area of the face image. The training process was carried out using four deep learning benchmarks model, which were DenseNet201, MobileNetV2, ResNet101, and EfficientNetb0. Besides that, deep learning has a very important number to determine that is learning rate, which impact the deep learning results. There were five learning rates, which were 10-1, 10-2, 10-3, 10-4, and 10-5. Learning rate values were then compared with four deep models learning to obtain high accuracy results in a short time and simple algorithm. Finally, the testing process was carried out on test data using a deep learning benchmark model in accordance with the training process. The research results showed that a learning rate of 10-2 from the MobileNetV2 method produced an optimal performance with a training validation accuracy of 99.60% within a time of 312 min and 28 s.