Segmentation Of Foreground Targets Based On Automatic Markers In Color Image

With the rapid development of information technology, image processing technology plays an increasingly important role in some fields such as military defense, medical diagnosis, weather monitoring and so on. As a basic computer vision technology, image segmentation is one of the key technology in image processing, and the segmentation results have a direct impact on the performance of the image analysis. Existing image segmentation methods have some limitations in the fields of practical applications, such as inaccurate segmentation of targets, time consuming, manual intervention and so on, therefore image segmentation has been one of problem in the computer vision technology and the researched hot spot of experts and scholars at home and abroad in recent years.This paper studies a segmentation method of targets in color images, whose aim is to extract the foreground targets of color images through an effective segmentation method. This article describes the basic theories and common methods of color image segmentation and proposes a segmentation method with block units of pixels, whose speed of segmentation increases dramatically compared to traditional segmentation method of pixels domain. It also introduces a learning mechanism with "soft segmentation" instead of the traditional "hard segmentation" to improve the accuracy of segmentation and puts forward one growth algorithm of constrained minimum spanning tree to solve the problem of imprecise segmentation with the traditional growing algorithm of minimum spanning tree.This paper mainly consists of the extraction of block feature, the auto-marking of seeds and regional growth, whose specific methods are as follows:the image set is divided into a training set and a test set. At the extraction of block feature stage, every image is blocked, each block is one set of some pixels, color moments and DCT coefficient of blocks are extracted as block characteristics. At the extraction of block feature stage, each feature vector of the training set is marked as category (foreground or background) and trained using support vector machine, so every feature vector in a test set could be marked by the training model, then filtering the outliers which are marked inaccurately through the three layer monitoring model of the outliers. At regional growth stage, the pixels blocks which have been marked foreground targets at previous stage are regarded as root nodes to construct constrained minimum spanning tree, followed by the merging algorithm of constrained minimum spanning tree to achieve the growth of the region and the final segmentation.This segmentation method is validated in Corel datasets in the paper. It’s proved that the method has high performance and can be applied in different types of images which won’t be disturbed by domain knowledge, which has certain theoretical guiding significance and engineered application value.