Research Of Related Image Segmentation Algorithms Based On Random Walks

Image segmentation as a component of image processing plays a very important role of computer vision. Image segmentation technique extracts interested target from the input image, has been widely used in artificial intelligence area, such as image analysis, image retrieval, and image recognition. Since there exist some complex color or texture features of natural images, it's hard to get the accurate segmentation result. Random walks(RW) algorithm which based on graph is a semi-automated segmentation method, has a very good effective when segment images that have noise, weakness boundary or even lack of boundary. Meanwhile, RW has advantages of segmentation speed, doesn't need iterate, can be easily extend to multi-dimensional space as well. Random walks with restart(RWR) adds a restart probability which return to the start point.However, the current restart probability of RWR mainly set based on the empirical value, and the algorithm also are sensitive to the seed quantity or placement. In this paper, we study the RW theory and image segmentation technique thoroughly, propose an adaptive restart random walks algorithm which can be used for interactive image segmentation. The concrete works of this paper as below:1. Study the current image segmentation techniques, classify and introduce the methods, also compare their advantages and disadvantages.2. Analyze several interactive image segmentation methods which based on Graph-cut, edge and region. At the same time, study random walks theory and random walks with restart theory.3. Modify the restart probability of random walks with restart, use PCNN edge detection method and substitute the presetting value with adaptive parameters.4. Combine with generative learning algorithm, change the start point from the non-marked pixel to markup pixel, and build generative model for each label such that the follow-up added labels don't have influence on the previous seed labels, this also prevent the repeated calculation and improve the execution speed.5. Improve the similarity measurement between pixel and label. We use the average of all the steady-state probabilities between pixel and the seeds with same label to replace the first arrival probability. Meanwhile, we also utilize the interior information of labels to guide the algorithm to obtain better performance during process the weak-boundary and complicated texture images.