Facial Expression Recognition in the Presence of Partially Occluded Images Using Higher Order Spectra

Facial expression recognition (FER) still be ongoing research and considered as a challenging task largely because of uncontrolled environmental conditions, for example, the presence of partially occluded images (masks or sunglasses) on the human face. To address the issue, this paper proposed a new approach of using higher order spectra (HOS) to improve recognition performance of FER under partial occlusions. HOS or second-order statistics the so-called bispectrum is used to reconstruct the occluded texture feature based on the configuration and visual properties of the human face. The bispectrum could capture a contour (shape) and texture information of the facial emotion whose enable the effective implementation algorithm in FER. In this framework, first the 2D facial spatial images are projected into 1D signal by means of Radon transform. The Radon transform has properties of rotation and translational invariants; thus, it can preserve any variation in pixel intensities. Then, the projected 1D signal was analysed using HOS to obtain bispectrum magnitude plot whose exhibit the behaviour of different emotions. A set of bispectral statistic features were extracted from the bispectrum plot and used as informative features to recognize the emotions. Linear discriminant analysis (LDA) was adopted to reduce the data features before fed as input to Support Vector Machines. A series of experiments have been conducted on CK database. The obtained results show that the recognition rates of occlusion of the upper face give the accuracy of 93.1%; thus, it is promising.