Optimized PHOG and KNN for Robust Frontal Facial Expression Recognition against Occlusion

Abstract

This paper presents an efficient for emotion recognition under occlusion mode of frontal facial images. The proposed algorithm firstly uses combination of Viola-Jones algorithm with skin color information for pure face detection. Then, in the detected face region, the proposed algorithm extracts an optimized Pyramid Histogram of Oriented Gradient (PHOG) descriptor that includes 4 pyramid levels and 4 bins for histogram. The proposed algorithm finally uses a KNN (K Nearest Neighbor) multi-classifier with Euclidean distance which results in high recognitions rate over large databases. The experiments over the RaFD face database show that the average recognition rate of the proposed algorithm for detecting seven common emotions (happy, sad, disgust, fear, angry, surprise and neutral emotions) is 99.52% for non-occluded condition. Moreover, the average recognition rates of the proposed algorithm on JAFFE and CK+, which are another popular face databases, are more than 94.6% and 99.1%, respectively. Averagely, the proposed algorithm achieves to 97.73% emotion recognition rate on over about 2000 facial images. On the other hand, when only half of face image is considered as the system input, the average recognition rate achieves to more than 88.57%, 85.5%, and 96.7% over RaFD, JAFEE and CK+ face images, respectively. Since the proposed algorithm shows high robustness against 50% occlusion in input face images, this algorithm can be used in occlusion conditions too.