Radial mesh pattern: a handcrafted feature descriptor for facial expression recognition

Abstract

Facial expressions are extremely important in the social interaction as they can display the internal emotions and intentions of an individual. Accurately classifying the facial expressions into various categories is the main task in Automatic Facial Expression Recognition (AFER) systems. The existing local based techniques, at times suffer and generate same feature values for different image portions such as edge, corner and flat regions. To address this issue, Radial Mesh Pattern (RMP), a local texture based approach based on the chess game rules is proposed. With reference to the center pixel in a 5 × 5 neighborhood, the possible positions of Rook, Bishop and Knight are determined and based on these positions, the features are extracted. In this paper, not only binary weights, but also other weights such as fibonacci, prime, natural, squares, odd and even weights have been utilized for feature extraction. To validate the efficiency of the proposed method, RMP is implemented on six 'in the lab' datasets. The performance is measured through recognition accuracy and the results obtained from experiments demonstrate the efficiency of RMP over standard existing methods.