Loading...

Intensity Estimation of Facial Action Units Utilizing Their Sparsity Properties

Mohammadi, Mohammad Reza | 2015

527 Viewed
  1. Type of Document: Ph.D. Dissertation
  2. Language: Farsi
  3. Document No: 47478 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Fatemizadeh, Emad; Mahoor, Mohammad Hossein
  7. Abstract:
  8. The most popular system for quantification of the facial behaviors and expressions is the Facial Action Coding System (FACS). FACS provides a description of all possible and visually detectable facial variations in terms of 33 Action Units (AUs). The activation of each AU leads to a slight variation in the facial appearance, and any facial expression can be modeled by a single AU or a combination of AUs. Definition of AUs is such that they are sparse in multiple domains. The goal of this dissertation is utilizing these sparsity properties to develop an effective algorithm for automatic intensity estimation of AUs. One of the sparsity domains of AUs is the spatial domain that means the impression region for each AU is a small part of the face. We use this property to develop an algorithm for removing the subject-dependent information from the extracted features from the sequence frames of each subject. This algorithm, as a preprocessing, enhances the quality of the features and improves the results of the next stages. Co-occurrence is the next sparse domain for AUs. In other words, the number of active AUs in any frame is very limited. Based on this property, we propose a supervised dictionary learning algorithm to estimate the intensity of AUs simultaneously. In the optimization problem of the proposed supervised dictionary learning we utilize other properties of AUs including their non-negative intensity that results in more suitable dictionary. The third sparsity domain is time that means each of the AUs are inactive (with 0 intensity) in most of the frames. We use this property and some other statistical characteristics of AUs to develop a novel algorithm based on the source separation idea. In this algorithm, each AU is considered as a source, and the intensities of AUs are estimated in a semi-blind source separation fashion. We use the unlabeled test data in this algorithm to obtain a subjectadaptive method. The ICC measure for this algorithm is obtained 0.73 that is better than state-of-the-art algorithms
  9. Keywords:
  10. Sparse Representation ; Dictionary Learning ; Facial Action Coding System ; Blind Sources Separation (BSS) ; Content-style Decomposition

 Digital Object List

 Bookmark

No TOC