Facial Expression Recognition Based On Shearlet Transform

Facial expression recognition arouses interest of research community owing to its great prospect in the applications of human-computer interaction and intelligent transportation systems. Facial expression recognition is mainly involved with two issues, one of them is how to get the features of the facial expression, the other one is how to carry out the analysis of the expression classification. This thesis mainly makes researches on the two issues mentioned above and put Shearlet Transform used in the state of the human facial expression recognition. As a new multi-scale geometric analysis method, The Shearlet transform not only has multi-resolution characteristics of wavelet transform and time-frequency localization properties, and also has a strong directional sensitivity and anisotropy. The main work and innovation of this thesis are organized as follows:1. The properties of the two-dimensional separable shearlet transform (2D-DSST) are researched in the field of image processing, and establish the regulation of how to extract the expression features from the facial images at shearlet domain based on this. Further more, it makes instructions on the selection of the facial expression database and pre-processing method, then changes the dimensions with the method of nearest neighbor interpolation to meet2D-DSST requirements.2. A method based on DSST and Support Vector Machine (SVM) for Facial expression recognition is proposed. In this method, the low and certain high frequency component in DSST are extracted and combined as features. This thesis also quantitatively analyzes the relationship between the recognizing results and the scale as well as each high-frequency component. The experimental results show that high-frequency component is helpful for recognition, but the recognition rate is low. Combining low frequency with high frequency component can effectively extract the essential characteristics of the expression, and improve the recognizing results. It is verified the proposed algorithm through comparing with other methods.3. A method for selecting an optimized set of the scale and direction of DSST is investigated. This thesis proposes the using of a separability judgment to evaluate the separability of different scales and directions. Then we only use those scales and directions that can better separate different expressions, in order to reduce the dimensions of the features and computation, and to improve the efficiency of the system.4. A method based on DSST and Compressed Sensing (CS) for Facial expression recognition is proposed. The compressed sensing theory is introduced. Using of the characteristic of the shearlet coefficients is very sparse, the approach of the facial expression recognition based shearlet and the compressed sensing is proposed. Compare the results with recognition algorithm based on DSST and SVM, in case of small sample it obtains the approximate recognition rate, while the recognition rate improves signifcantly in case of large sample. Because it does not contain empirical parameters and needn’t set parameters, so it is convenient compared with SVM, and it also solutions effectively the problem of SVM classifier to be ineffective for large sample.