Emotion recognition techniques with rule based and machine learning approaches
Authors:
Aasma Aslam,
Babar Hussian
Abstract:
Emotion recognition using digital image processing is a multifarious task because facial emotions depend on warped facial features as well as on gender, age, and culture. Furthermore, there are several factors such as varied illumination and intricate settings that increase complexity in facial emotion recognition. In this paper, we used four salient facial features, Eyebrows, Mouth opening, Mouth…
▽ More
Emotion recognition using digital image processing is a multifarious task because facial emotions depend on warped facial features as well as on gender, age, and culture. Furthermore, there are several factors such as varied illumination and intricate settings that increase complexity in facial emotion recognition. In this paper, we used four salient facial features, Eyebrows, Mouth opening, Mouth corners, and Forehead wrinkles to identifying emotions from normal, occluded and partially-occluded images. We have employed rule-based approach and developed new methods to extract aforementioned facial features similar to local bit patterns using novel techniques. We propose new methods to detect eye location, eyebrow contraction, and mouth corners. For eye detection, the proposed methods are Enhancement of Cr Red (ECrR) and Suppression of Cr Blue (SCrB) which results in 98% accuracy. Additionally, for eyebrow contraction detection, we propose two techniques (1) Morphological Gradient Image Intensity (MGII) and (2) Degree of Curvature Line (DCL). Additionally, we present a new method for mouth corners detection. For classification purpose, we use an individual classifier, majority voting (MV) and weighted majority voting (WMV) methods which mimic Human Emotions Sensitivity (HES). These methods are straightforward to implement, improve the accuracy of results, and work best for emotion recognition using partially occluded images. It is ascertained from the results that our method outperforms previous approaches. Overall accuracy rates are around 94%. The processing time on one image using processor core i5 is ~0.12 sec.
△ Less
Submitted 28 February, 2021;
originally announced March 2021.
