Research On Facial Expression Region Extraction Method Based On Graph Cuts Optimization

Traditional Graph Cuts model usually uses color features and texture features to build the energy function of image segmentation. However, the edge of facial expression region is both weak and blurry, the color and texture information in this region is not so abundant. So, if we use traditional Graph Cuts model to extract facial expression region, the result is always undesirable. In details, the main research achievements of this paper can be described as follows.Firstly, we studied variational functional model and Graph Cuts optimization model in image segmentation field, by using the Cauchy-Crofton formula in Integral Geometry, we can easily convert variational functional model to Graph Cuts optimization model. Then, we studied how Active Shape model work. After this, with the help of open-source software STASM supplied by Milborrow, we experimented on the landmarks extraction of facial expression image. The results showed that by using ASM, we can get the key points of expression region quickly and precisely.Secondly, we proposed a Graph Cuts optimization model based on the contour band of expression region. By joining landmarks around the boundary of expression region, we can get a closed curve that approximatively describe the border of expression region. Then, by dilating and corroding the closed curve, we get the contour band that contain the real boundary of expression region. The external profile can be viewed as background seeds and the internal profile can be viewed as foreground seeds, then we can add these seeds to the graph structure that corresponding the energy function of segmentation. Another strategy of how to use the contour band is inspired by GCBAC method, we build the graph structure only in band region, then we change the way that setting weight of t-link and n-link and simplify the graph structure.Finally, based on the idea that the shape of closed curve should be a shape prior of expression region, we proposed a Graph Cuts optimization model based on the shape distance function. The distance from pixel to the boundary of shape prior can be viewed as a penalty, and we add this penalty both in data item and smooth item of the energy function.To solve the extreme value problem of energy function, we used the classical maxflow/mincut algorithm in graph theory analysis. A lot of experimental results demonstrate the advantages of our proposed method in terms of robustness and accuracy.