Interactive Image Segmentation Based On Multiple Instance Learning

Image segmentation attracts many researchers attention in the field of computer vision. It is the basic issue but also the difficult issue in image processing algorithms. Many scholars have been involved in research in this field and made many new algorithms. But still not can be applied to all situations, universally applicable image segmentation method. Image segmentation is the basis of image processing, and image processing is a very wide range of fields. It includes image recognition, image understanding, image enhancement, image classification, and so on.This paper is study the interactive image segmentation problems. According to the image and bounding box provided by users, then the algorithm should find the exact location of the target in the box. We make full use of implicit information of the bounding box and images, such as the continuity of the target, the compactness of bounding box, the Markov property of regions. After analyses of image segmentation problems and learn from previous studies, we propose two segmentation algorithm is as follows:1. We propose a method based on boosting and multiple instance learning framework, which is naturally transformed a image segmentation problem into a multiple instance learning problem. We propose a method to build bags and instances which is feasible and reliable. Then we combine the bags into a group randomly, each group bags consists of some positive bags and some negative bags, trans a week classifier from each group. finally, we assign specific weights of these week classifiers by boosting method to form into a robust classifier. Experimental result shows that segmentation result by this classifier can achieves sufficiently accurate.2. We also make full use of the image hidden information, present a new segmentation algorithm based on Markov random field, which method contains both of structure information and other appearance information. We consider the label of each pixel as a random variable and view the assign label process as a probability event. Then the label of the whole image will be a random field. We make some assumptions with the random field make it to be a Markov random field. Finally, we use Markov random field theory and the multiple instance learning theory optimize image labels, and update it in each iteration. It will achieve the results of image segmentation.