Study On Class-specific Fragment Based Top-down Image Object Segmentation

As a kind of basic technology in image engineering, image segmentation is crucial to many fields such as Machine Vision and Object-Oriented Multimedia technology. Image segmentation has broad applications in nearly all areas related to image. Image object segmentation which is to extract the interesting object from its clutter background in an image is the highest goal of image segmentation. Therefore, the study on the theories and related technologies of image object segmentation is quite meaningful.Aiming at extracting the interesting object in images, this thesis analyses and summarizes some related theories and methods. In this thesis, we first compartmentalize the goal of image segmentation based on the analysis of its definition. Through the wide study on image segmentation methods, we find that the top-down segment process which utilizes the prior knowledge of the object is an effective method of image object segmentation.According to the above conception, we investigate an image segmentation method based on class-specific fragment, called CSF-SEG for short, which belongs to the top-down processes. In this method, the construction of an object by fragments is somewhat similar to the assembly of a jigsaw puzzle, where we try to put together a set of pieces such that their templates form an image similar to a given example. However, there are some problems in the original method, such as low availability of the extracted fragments, error position of the match results, and high complexity of the algorithm etc. In this thesis, a novel method to extract fragments is proposed, and it improves the availability of the extracted fragments; a novel match method called HER is also proposed, which is a Hybrid one combining the object's Edge and Region features, and its advancement is showed in the experimental results; a prototypal top-down image object segmentation system named IO-SEG is devised and implemented in this thesis, which is based on class-specific fragment and integrates the methods proposed in this thesis.