Salient feature and reliable classifier selection for facial expression classification

A component based approach improves classification of discrete facial expressions over a holistic approach

“…For the Cohn-Kanade dataset, the studies which employed six expression classes (happy, sad, surprised, fearful, angry and disgusted) achieved 93.85% [30] and 93.8% [56]; and those with seven expression classes (the typical six plus neutral) achieved 93.8% [17] and 93.3% [13]. For the MMI dataset, the studies which employed six expression classes achieved 82.68% [8] and 92.6% [12]; and those with seven expression classes achieved 88.9% [12] and 93.61% [21]. In comparison, our best results employing six expression classes achieved 80.83% via a component approach, and for a seven expression class we achieved 98.33% via a component action approach.…”

“…Happy, sad, surprised, fearful, angry and disgusted are the typical expression classes studied ( [6], [7], [8]), and neutral is used as a reference image in a dynamic analysis (aka holistic action/component action) ( [9], [10], [11], [12], [13], [14]) and as an individual class in a static analysis (aka holistic/component) ( [6], [7], [15], [16], [17], [18], [19], [20], [21]). Paul Ekman and Wallace V. Friesen hypothesise these expressions are appropriate to be universally recognized ( [22], [23]); and have devised a widely used system of metrics known as 'The Facial Action Coding System' (FACS) that captures facial actions that are applicable to facial expressions [24].…”

“…Edge detection ( [36], [37]) and histogram equalisation ( [20], [21], [31]) are commonly used to counter the effects of noise from poor lighting. The method of Local Binary Patterns (LBP) has also become popular for tackling poor lighting and has shown promising results in facial expression classification ( [12], [16]).…”

“…Wavelet compression techiques have been used to reduce noise by taking into account the spatial information of the face. For instance, Haar-like (using the Integral Image [38]) and Gabor wavelets have been used in [30] and ( [7], [16], [17], [18], [21]), respectively. Further studies involve the use of Active Appearance Modeling (AAM) [6] to describe the face using feature points ( [11], [14], [39]).…”

Facial Expression Recognition Using Facial Movement Features

“…These approaches are selected because they produced the stateof-the-art performance using a similar testing strategy and the same databases. In [40], only the recognition results of the leave-one-subject-out strategy is used here as this strategy is more similar to our leave-one-set-out cross validations. The recognition results in [41] were obtained by removing two JAFFE images named "KR.SR3.79" and "NA.SU1.79".…”