Research Of Natural Image Segmentation Algorithm

Image segmentation has long been a challenging problem in computer vision and pattern recognition. Its aim is to find regions in an image which are coherent in semantic. To reach this, a large number of various methods have so far been proposed. We can broadly categorize them in two directions:pixel-wise and region-wise.Compared with the pixels, regions from an over-segmentation of the original image can contain more information. By transforming an image with millions of pixels into a few initial regions, our CCTA based region merging method benefits from this reduction in complexity. The initial segmented image is represented by a region adjacency graph (RAG)--each node is one homogenous region, each pair of nodes is connected by the undirected edge, and the edge weight measures the similarity between adjacent regions. By Kruskal’s minimum spanning tree (KMST) algorithm, the adjacent consistent regions are iteratively merged so as to reach a global optimality.Recently, pixel-wise based active contour models (ACM) have been widely studied for medical image segmentation. The main idea is the level-set based surface evolution theory. These models can be classified as edge-based or region-based, and Chan-Vese (CV) is a most popular region-based model. In particular, CV model requires no image gradient information, and thus can segment objects with weak boundaries or without boundaries. However, it works poor for cases with the low-level inhomogeneity, e.g., natural images. By embedding the PIF (pixel inhomogeneity factor) proposed in CCTA, we develop a new active contour model that can achieve better results for natural images.Finally, extensive experiments and evaluation are conducted on a wide variety of natural images. Results show that our proposed two methods can reliably segment many objects out from their surrounding backgrounds.