Combining local descriptors and classification methods for human emotion recognition.

Abstract

Human Emotion Recognition occupies a very important place in artificial intelligence and has several applications, such as emotionally intelligent robots, driver fatigue monitoring, mood prediction, and many others. Facial Expression Recognition (FER) systems can recognize human emotions by extracting face image features and classifying them as one of several prototypic emotions. Local descriptors are good at encoding micro-patterns and capturing their distribution in a sub-region of an image. Moreover, dividing the face into sub-regions introduces information about micro-pattern locations, essential for developing robust facial expression features. Hence, local descriptors’ efficiencies depend heavily on parameters such as the sub-region size and histogram length. However, the extraction parameters are seldom optimized in existing approaches. This dissertation reviews several local descriptors and classifiers, and experiments are conducted to improve the robustness and accuracy of existing FER methods. A study of the Histogram of Oriented Gradients (HOG) descriptor inspires this research to propose a new face registration algorithm. The approach uses contrast-limited histogram equalization to enhance the image, followed by binary thresholding and blob detection operations to rotate the face upright. Additionally, this research proposes a new method for optimized FER. The main idea behind the approach is to optimize the calculation of feature vectors by varying the extraction parameter values, producing several feature sets. The best extraction parameter values are selected by evaluating the classification performances of each feature set. The proposed approach is also implemented using different combinations of local descriptors and classification methods under the same experimental conditions. The results reveal that the proposed methods produced a better performance than what was reported in previous studies. Furthermore, the results showed an improvement of up to 2% compared with the performance achieved in previous works. The results showed that HOG was the most effective local descriptor, while Support Vector Machines (SVM) and Multi-Layer Perceptron (MLP) were the best classifiers. Hence, the best combinations were HOG+SVM and HOG+MLP.