Log-Normal and Log-Gabor Descriptors for Expressive Events Detection and Facial Features Segmentation

Abstract

The current paper investigates the merits of the Log-Normal and Log-Gabor filters for the dynamic analysis and segmentation of facial behavior during facial expression sequences. First, a spatial filtering method based on the Log-Normal filters is introduced for the holistic processing of the face towards the automatic segmentation of consecutive "emotional segments" in video sequences. Secondly, a filtering-based method based on the Log-Gabor filters is applied as a feature-based processing for the automatic and accurate segmentation of the transient facial features (such as nasal root wrinkles and nasolabial furrows) and a precise estimation of their orientation in a single pass. We compared heuristic and machine learning based methods to evaluate the efficiency of the used descriptors for each task. When tested for automatic detection of "emotional segments" in 137 video sequences from the MMI, the Hammal-Caplier facial expression databases, and 20 recorded video sequences of consecutive appearance of multiple facial expressions, the proposed Log-Normal based descriptors achieved an accuracy of 89% with a mean frame error of 8 frames using a heuristic based processing. Higher performances were obtained using the SVM based method leading to an accuracy of 94% with a mean frame error detection of 3.1 frames. Tested on more than 3280 images from 5 benchmark databases (i.e. the Cohn-Kanade database, the CAFE database, the STOIC database, the MMI database, and the Hammal-Caplier database) the proposed Log-Gabor based descriptors for transient facial features detection achieved a mean performance of 82% using a heuristic based processing and a mean performance of 96% using the SVM based classification. The proposed method for the estimation of the corresponding orientation leads to an error of 2.7°.