Human Emotion Recognition in Video Using Subtraction Pre-Processing

Abstract

In this paper, we describe a new image pre-processing method, which can show features or important information clearly. Deep learning methods have grown rapidly in the last ten years and have better performance than the traditional machine learning methods in many domains. Deep learning shows its powerful ability particular in difficult multi-classes classification challenges. Video Facial expression recognition is one of the most popular classification topics and will become essential in robotics and auto-motion fields. The new system presented is a combination of new video pre-processing and Convolutional Neural Network (CNN). The new pre-processing method is proposed because we believe individual emotions are dynamic, which means the change of the face is the key feature. RAVDESS is the video set used, to train and test the neural network. From RAVDESS dataset the video songs without audio are taken for focusing on video frames differences. The chosen video set has six different classes of emotions. Each video presents a sentence in a melodious way. Based on the chosen video set, the new system with a new pre-processing method has been designed and trained. Later, the classification result of the new method has been compared with others in which the same dataset for video emotion recognition was used.