A Figure Extraction and Synthesis System by Learning Vector Quantization Neural Networks

Extracting complete figures from videos with complicated environments is difficult. A new figure extraction and synthesis system with capability of extracting figures from consecutive frames in a messy environment is proposed in this paper. A figure template is constructed based on the face detection results and some image processing techniques. Figural and non-figural features are extracted from the figure images. By means of these features, a learning vector quantization neural network (LVQNN) is applied to classify the uncertain regions into figural and non-figural objects. The extracted figure can be further synthesized into an optional cinestrip. Experimental results showed the proposed method successfully extract the figure object from a complex background environment.