Texture vs. multiresolution analysis of facial expressions: application to emotion recognition

摘要

The work presented here portrays a comparative facial expression analysis using texture and multiresolution approaches for automatic emotion recognition. The emotions can be recognised by taking note of the variations in spatial arrangement and intensity of the pixels, corresponding to the features being used for emotion detection in human interactions. Extensive texture and multiresolution analysis has been performed, and impact of noise, illumination, shift and scale changes in test images is discussed. For multiresolution analysis (MRA), we have used wavelet and curvelet algorithms. The experiments are performed over three databases viz. Cohn-Kanade, JAFFE and in-house database, and the global accuracies are given in terms of AUC of RoC, precision, recall, F-measure, etc., on five different classifiers namely SVM, MLP, K-NN, K~* and meta multiclass. We have bench-marking results under the noise and illumination-change conditions. A comparative performance is also given for texture and multiresolution analysis over all three databases. The outcome of evaluation and comparison indicates that MRA outperforms the texture analysis.