Article Preview
TopIntroduction
Facial expression recognition is an important area of Artificial Intelligent (AI) systems that falling in the domain of Affective Computing. The human facial expressions maintain a good correlation with human emotions but determining the proper expression class from a face image in different conditions like changing lighting, viewing angle, resolution, and different subjects is a thought-provoking problem. Also, the human facial expressions differ according to their race, religion and cultural background and the overlapping expression classes increase the possibility of miss-classification. As an example, the ‘fear’ expression is often misinterpreted with ‘Anger’ expression.
Automated facial expression recognition has many applications in different industries and such as Human-Computer Interaction (HCI), Companion Robotics, Pain monitoring, Child care systems, Driving monitoring system and Attentiveness Analysis, etc(Li, S., & Deng, W. 2020).This type of systems is categorized into two different approaches with one being Facial Action Coding (FAC) based and another based on geometrical and texture analysis of whole face Ekman, (P., Friesen, et al. 2002). The Facial Action Coding (FAC) approachtaxonomize the facial expressions as a combination of different Action Units (AUs) which are performed by a subject by activating different facial muscles Ekman, (Friesen, E., & Ekman, P. 1978). In contrast to that the holistic approach considers the face as a single unit and analyses the geometric and texture properties of the face without bothering as to which AU is performed (Shan, C., et al. 2009). The holistic analysis of facial expression has the advantages as it avoids the AU detection error and also intra-component variations can be modeled in this approach thereby enhancing the overall accuracy of the recognizer system.
A geometric based facial expression recognition system has three phase of computation 1. Facial landmark generation, 2. Extraction of features from landmark points and, 3. Classification. This study mainly focuses on feature extraction from facial landmarks points for better classification of expressions. The authors in the present scope have used the pre-trained Active Appearance Model (AAM) for face recognition for determining the geometrical location of facial components (Sagonas, C., et al. 2016). The AAM is an evolved version of the Active Shape Model (ASM) which is a statistical object recognition model and uses the shape and texture properties of an image Cootes, (Cootes, T. F., & Taylor, C. J. 1992). The AAM describes a face using 68 landmark points plotted over eyes, eyebrows, nose, lips and outer shape of the face. For designing an effective facial expression recognition system all the 68 landmark points are not necessary. For example, the landmark points around the lip region are more sensitive to expressions rather than landmark points around the nose. To remove comparatively irrelevant points out of those 68 landmark points some sort of algorithm is necessary. The authors selected the salientlandmark points by training a set of 68 MLP classifier that with the textured pattern of the neighboring pixels of thoselandmark points. The texture pattern is determined by computing the Histogram oriented Gradients (HoG) of the neighborhood of that landmark point. Now top performing 25 MLP out of 68 MLP are selected as indicator of salient landmark points. In the next stage of computation, the salient landmark points selected in this manner are used to form a triangulation set. The shape feature of each triangle in the triangulation sets is computed and reshaped as a matrix termed as Shape Information Matrix (SIM). The final feature for training is computed by extracting the HoG features of the SIM matrix. To determine the final expression class a final MLP classifier is trained with this feature set.