Facial emotion recognition under partial occlusion using Empirical Mode Decomposition

One of the challenges in automatic facial emotion recognition nowadays is the ability to handle with complicated environment conditions such as in the presence of partial occlusions of facial images. To address this issue, therefore this paper proposed to investigate the effect of facial emotion recognition in the presence of partially occluded images using empirical mode decomposition (EMD). EMD a multi-resolution technique which is adaptively decomposed non-stationary and nonlinear data into a small set of frequency component known as intrinsic mode functions (IMFs). In this work, the face image is firstly projected into 1D signal using the Radon transform. The projected 1D signal is subjected to EMD to extract the significant features based on IMFs. The obtained IMFs features are further reduced using PCA plus LDA to reduce the dimension of the features. Then, the reduced feature vector is used as input to Support Vector Machines (SVM) classifier for recognizing seven facial emotions. A series of experiments has been conducted on the CK database under four different modes of occlusion such as right face occlusion, left face occlusion, upper face occlusion and lower face occlusion. The experimental results show that the upper face occlusion contributes the highest recognition rate which is 93.91%, thus the proposed method demonstrates the promising results.