Automatic Semantic Video Object Extraction

This is a joint work with Dr. Jianping fan. from the Dept. of Computer Science, University of North Carolina at Charlotte. Our works on this topic focus on color edge detection, seeded region growing and seeded region aggregation for salient object generation. Salient objects are needed to be the semantic objects from human point of view, but they should have the capability to indicate the presence or absence of the associated semantic video concepts. Image segmentation and analysis is the first step for building multimedia database systems: feature extraction. Now, we are still working on this topic to find more efficient and effective ways to obtain salient objects and their representative visual features. As shown in the following figures, the color edges are first obtained by using our color detection algorithm, and these connected color edges are merged and labeled by the same symbol, the centers of these neighboring connected color edges are taken as the initial seeds for automatic seeded region growing.

Figure 1. Content-based image analysis

For more information about this part, please visit Dr. Jianping Fan's Home page: http://www.cs.uncc.edu/~jfan

Towards facial feature locating and verification for omni-face detection in video/images

Face detection is important in image/video content analysis and organization since the most important object in those medias is often human being. We propose a facial feature based omni-face detection algorithm in this paper. While utilizing the skin color model for face cue detection, a binary skin region refinement strategy is applied to eliminate the error introduced by skin model. Then, a region based adaptive threshold selection scheme is employed for facial feature segmentation. After the facial feature filtering, an orientation, pose and scale invariant face verification strategy is used to verify the detected candidate face region. Experimental results demonstrate successful detection over a wide variety of facial variation in color, background, view and orientation from different types of video collections.

Figure 2. Omni-face detection

Figure 3. Binary skin region refinement (with image from left to right represent the original image, skin color segmentation result, layer 1 and layer 2 respectively).

Figure 4. Facial feature extraction (with image from left top to right down represent the original image (I), Prewitt edge of the image (EI), segmented face candidate region (A), facial feature segmentation result and filtered facial feature areas (MRB) respectively )

Figure 5. Face detection results