Continuous learning of a multilayered network topology in a video camera network
Journal on Image and Video Processing - Special issue on video-based modeling, analysis, and recognition of human motion
| Hi-index | 0.01 |
Face recognition based on video has received significant attention in the past a few years. However, the acquainted facial images in video from a distance are usually small and the quality of them is low. Enhancing low-resolution (LR) facial images from the video sequence is of importance for performing face recognition. Super-resolution (SR) reconstruction is one of the most difficult and ill-posed problems due to the demand of accurate alignments between multiple images and multiple solutions for a given set of images. In particular, human face is much more complex compared to other objects which are addressed by the majority of the super-resolution literature. Super-resolution from facial images may suffer from subtle facial expression variation, non-rigid complex motion model, visibility and occlusion, and illumination and reflectance variations. In this dissertation, the objective is to design algorithms to tackle the problems brought by the special characteristics of human face. The techniques we present in this dissertation address facial expression variations, non-rigidity of face and illumination changes as follows: (1) Closed-loop Construction of Super-resolved 3D Facial Texture in Video. A closed-loop system for incremental super-resolution of video and its use for face recognition. The system uses a generic 3D model of the face and compensates for changing illumination and 3D pose in video. (2) Super-resolve LR Facial Images by Treating Face Images Non-uniformly. A super-resolution system that explicitly accounts for facial expressions by treating the face as the composition of local face entities (eyes, nose, mouth, eyebrows and rest of the face) and performing appropriate distortions to the face. (3) Register LR Facial Images Using Global Parametric Model and Free Form Deformation. We propose a new method for enhancing the resolution of low-resolution (LR) facial image by handling the facial image in a non-rigid way: firstly a global model is employed to track the faces through the video sequence; then a B-spline based Resolution Aware Incremental Free Form Deformation (RAIFFD) model is used to recover a dense local non-rigid flow field. (4) Evolutionary Feature Synthesis for Facial Expression Recognition. A genetically-inspired learning method for facial expression recognition. Unlike current research on facial expression recognition that generally selects visually meaningful feature by hands, our learning method can discover the features automatically in a genetic programming-based approach.