Emotion detection using 2-dimensional mode decomposition methods from facial images

People are in interaction and communication with themselves and with people around them. This communication occurs not only verbally but also through gestures and facial expressions. Body movements and facial expressions, whether voluntary or involuntary, carry significant clues about an individual's emotional state. These universal behavioral patterns exhibit common characteristics across different cultures in reflecting emotions. The development of emotion recognition technology holds significant importance for society and the business world. In particular, in the retail and marketing sectors, understanding customers' emotions and preferences enables more effective presentation of products. Additionally, the use of emotion recognition systems is increasing in areas such as security and education. For example, in education, analyzing students' facial expressions can assess their interest in lessons, aiding in the improvement of learning processes. Studies on facial expression analysis and emotion recognition can be divided into two main categories: traditional methods and deep learning-based methods. While traditional methods perform classification by using feature engineering and machine learning methods, deep learning-based methods have come to the forefront, especially with the use of convolutional neural networks. In this thesis, the advantages of using mode decomposition methods in solving the problem of emotion recognition from two-dimensional face images have been examined. Face images are represented by a finite number of modes, and these modes are used instead of face images in the proposed approach. Three important two-dimensional mode decomposition methods have been considered. The modes obtained by the Two-Dimensional Empirical Mode Decomposition (BEMD), Two-Dimensional Variable Mode Decomposition (2D-VMD) and Geometric Mode Decomposition (GMD) methods have been used to obtain the descriptive features that for emotion recognition. BEMD, 2D-VMD and GMD each have different advantages and challenges. While BEMD stands out in processing nonlinear and nonstationary signals, it may experience mode mixing. While 2D-VMD offers a clear advantage in isolating certain frequency bands, it may be limited in adapting to various data sets. Although GMD is extremely effective in processing geometric structures, its applicability in more general image processing scenarios may be limited. In solving the problem of emotion recognition from face images, the contributions of using the modes obtained with these three methods instead of face images were tested with both traditional machine learning methods and deep learning approaches, and the obtained emotion estimation results were evaluated.