Recognizing facial actions using gabor wavelets with neutral face average difference

Abstract

This paper describes a new pre-processing step to classify facial expression. Previous works suggest that Gabor wavelets applied to recognize facial expression images subtracted from neutral face from the same subject could achieve good recognition rate under controlled condition as eye and month alignment. We propose a recognition system where the Gabor kernels are applied on facial expression subtracted from a averaged neutral face. A fast pre-processing technique that generates a small dimension output data is also proposed. A correct recognition rate of 86.6% is obtained in a 7 upper face actions and 81.6% in a 7 lower face actions detection problem using a Neural Network based classifier. The performance is evaluated in a heterogeneous subject database with head motion and lighting variations.

Facial Action Unit Recognition by Exploiting Their Dynamic and Semantic Relationships

“…Among the holistic facial feature extraction methods, there are three major variations: motion extraction by dense flow estimation [3], [4], motion extraction based on difference images [4], [5], [6], [7], and single-image-based methods [21], [8].…”

“…Further operations are performed on the difference image to extract the image features including the Gabor-wavelet-based feature [6], [7] and low-dimensional principal component coefficients [4], [5], [7]. The difference-image-based methods have the advantage of being robust to the illumination and skin color variations but cannot extract the motion direction.…”

“…In the facial feature extraction stage, the previous work can be summarized into two major groups: holistic techniques [3], [4], [5], [6], [7], [8], where the face is processed as a whole, and local techniques [3], [4], [9], [10], [11], [12], [13], [14], [15], [16], in which only a set of specific facial features or facial regions are considered. The approaches on AU classification could be grouped into spatial approaches [17], [3], [4], [10], [7], [5], [6], [11], [12], [16], [8], [13], [18], where AUs are analyzed frame by frame, and spatial-temporal approaches [3], [9], [14], [15], [19], where AUs are recognized over time based on the temporal evolution of facial features. Basically, these methods differ in either the feature extraction techniques or the AU classification techniques, or both.…”

Machine Learning Methods for Social Signal Processing

“…For instance, (Lucey et al 2011) addressed the problem of pain detection by applying SVMs either directly to the image features or by applying a two-step approach, where AUs were first detected using SVMs, the outputs of which were then fused using the Logistic Regression model. Similarly, for the static classification of AUs, where the goal is to assign to each AU a binary label indicating the presence of an AU, the classifiers based on NN (Bazzo & Lamar 2004, Fasel & Luettin 2000, Ensemble Learning techniques (such as AdaBoost (Yang et al 2009a) and GentleBoost (Hamm et al 2011)), and SVM (Chew et al 2012, Bartlett et al 2006, are commonly employed. These static approaches are deemed context-insensitive as they focus on answering only one context question, i.e., how.…”

AutoBAP: Automatic Coding of Body Action and Posture Units from Wearable Sensors

“…Automatic implementations of FACS include a system trained to automatically detect action units in order to differentiate fake from real expressions of pain [10] and to analyse expressions of neuropsychiatric patients [11]. Techniques to achieve this include analysing permanent and transient facial features in frontal face image sequences [12], using independent component analysis and support vector machines [13] and using Gabor wavelets with neutral face average difference [14].…”