An analysis of machine learning techniques for multimodal recognition

This study delves into the realm of machine learning (ML). Recognizing the significance of ML in modality recognition, such as in Facial Expression Recognition (FER), Eye Movement Detection (EMD), and Voice Identification (VI), researchers have proposed varied ML techniques. However, existing research encounters constraints such as time and financial limitations, impeding a comprehensive comparison and selection of optimal ML techniques for modality recognition. To bridge this gap, our study conducts a preliminary investigation, comparing ML techniques based on their accuracy levels for modality recognition. Utilizing a Systematic Literature Review (SLR) approach, our study focuses on publications from the past five years. Specifically concentrating on ML applications in FER, EMD, and VI, we consider accuracy measures as key criteria. This study provides a comprehensive overview of ML techniques applied for FER, EMD, and VI. Notable achievements include Pure Convolutional Neural Network (CNN)'s exceptional accuracy on specific datasets, but limitations such as inadequate representation of facial expressions in some datasets are identified. Future research suggestions include sophisticated model testing and the use of Gaussian approaches for enhanced accuracy. Also, EMD highlights a variety of ML techniques achieving accuracy rates between 66% and 100%, with challenges prompting recommendations for advanced models, transfer learning, and addressing issues like sunglasses hindering facial feature detection. For VI, techniques like CNN, EDT, MLP, and others achieve accuracy rates spanning from 15% to 98.13%, with proposed solutions including dataset scaling, hybrid model creation, and combining EfficientNet with acoustic/linguistic knowledge to categorize a greater number of new words. In conclusion, this preliminary investigation lays the groundwork for optimizing ML techniques in modality recognition. The systematic comparison of various ML techniques and their accuracies provides valuable insights for researchers and practitioners. Identified limitations and proposed future research directions contribute to the ongoing evolution of ML applications across diverse domains.