The Reliability, Safety, and Control Principle on Recognition of Facial Expression Through Deep Learning to Support Responsible AI Implementation Policy

Abstract

Artificial Intelligence (AI) is a system that uses the power of science and technology to create a system that can think and act like humans. It is also a study of creating intelligent computers that can carry out tasks that typically need human intelligence. The use of AI often involves collecting vast amounts of personal data, and if this data is not adequately monitored and protected, it can lead to breaches of privacy and security as well as unethical. If AI is not addressed properly, it may lose out on the advantages and opportunities that these technologies may provide. In view of this, the Malaysian Government, through MOSTI, has launched National AI Roadmap 2021-2025, which also emphasizes Responsible AI practices, including the reliability, safety, and control of AI to minimize harm and challenges. In addition, with the shift from laboratory-controlled to demanding in the-wild circumstances for facial expression recognition (FER) and the relative success of deep Learning methods in multiple areas, deep neural networks are progressively being used to learn discriminative representations for automated FER. The latest deep FER systems mainly focus on two issues overfitting due to a lack of enough data and unrelated variations like lighting, head pose, and identity bias. In this paper, we propose a solution for facial expression recognition that integrates Convolutional Neural Network (CNN) and specific image pre-processing steps. CNN achieves better accuracy with given data. Yet, there are no available datasets with adequate data for facial expression recognition with deep architectures. Thus, the pre-processing technique is applied to extract only five expression-specific features from a face image detecting emotions such as happiness, surprise, sadness, anger, and normal. A sufficient number of Eigen face values are chosen. These Eigen faces can make up any image of training data when added with the right proportions; after all eigenvalues from each trained image are evaluated, a new input image can be processed by calculating the Euclidean Distance between the Eigenvalues of the input image and every training image within the help of MATLAB software the lowest distance will determine the current expression.