Research On Key Techniques Of Image Segmentation Based On Graph Cuts

With the continuous development of the computer vision and digital image processing technology, image segmentation technology has made a huge progress. In recent years, people put forward many segmentation methods based on specific theory due to the actual demand. As an image segmentation method based on graph theory, segmentation algorithm based on graph cuts transforms the segmentation problem into solving the minimum cut of the graph on the graph theory. Through the construction of the suitable energy function, the regional information and edge information of the image are effectively constrained. The corresponding graph can be built according to the energy function. The minimum cut of the graph and the global optimal solution of the function will be obtained by the max-flow/min-cut algorithm. Graph cuts have been widely used in the field of image segmentation. However, traditional segmentation based on graph cuts still exists some problems on segmentation accuracy and so on. In order to solve these problems, the paper designs an algorithm which combines affinity propagation based on adaptive weight feature and graph cuts. The main work includes the following two parts:On one hand, in the process of the improved affinity propagation, firstly, we consider that the general affinity propagation algorithm usually regards the pixels of the image as data points to construct the similarity matrix, which may cause large calculation. So the paper divide the image into many blocks with the same size and the blocks will be regarded as data points to construct the similarity matrix. It can effectively reduce the computational redundancy and improve the speed of clustering. Secondly, as the construction of the similarity matrix only uses the color information of the data points, which may cause low clustering quality, our paper proposes the adaptive weight feature. The color feature, texture feature and shape feature of the image are selected to build the feature space of the data points. The weight of each feature is assigned according to their distribution of the image automatically. Finally, the paper uses the feature information and the position information of the data points to measure the similarity and get the expected clustering regions.On the other hand, in the process of the construction of the energy function and the graph, firstly, we need to mark the clustering regions as initial labels and choose proper models to express the property of the clustering regions. Secondly, when constructing the data term of the energy function, the image data is mapped to the high dimension space implicitly for its complexity and the image data can be separated linearly. When constructing the smooth term of the energy function, we use the distance between the adjacent pixels to measure the discontinuity of the region edges in the image. At last, According to the energy function, corresponding weights are assigned to the edges of the graph. The max-flow/min-cut algorithm is utilized to search for the minimum cut of the weighted graph iteratively. In every iteration, the regional models will be updated until we get the final results.