Research On A Method Of Granular Cognition Based On Bottom Features Of Images For Object Extraction

The application domain of unsupervised discovery and segmentation of objects is quite broad. It includes video coding, video surveillance, image and video editing, image and video retrieval, human computer interaction, image understanding and object recognition, etc. According to the anaylsis of the numerous applications, it was found that there are three concerned issues:(1)the degree of the closeness, by which discovered objects approach the correct objects, such as the degree of agreement between edges and the capability of restraining possible interferences;(2) the generality and the extented capacity of the proposed algorithms—it is often impossible to make an algorithm discover or track objects perfectly at a time due to the fact that there is a great variety of images and videos, so it is necessary that the proposed algorithm have the capability for sustainable upgrade and good transplant so that the correct objects can be discovered and tracked in a closer and closer way;(3) the complexity—the proposed algorithms must be handy, better linear, so that these algorithms are able to be used to discover and track objects under condition of limited resources such as embedded devices and wireless sensor networks which are witnessing a rapid development. Aiming at solving the problems ibid, centering around the tracking of objects in videos and the unsupervised discovering of objects in images, and using psychological cognition model for reference, this dissertation set up a granular cognition platform for the discovering of objects. And then the platform is used in the unsupervised rough extracting of salient objects. The main work of this dissertation is as follows.This dissertation starts with the study of the algorithm for the tracking of objects moving in complex motion in videos. Three factors contribute to more complexity of motion:longer sampling period, an moving object with complex appearance and nonrestraint movement and occlusion. And then mean shift algorithm loses its target due to too low a Bhattacharyya coefficient. To treat it, mean shift algorithm is improved based on fast color thresholding and region merging in this dissertation. Visual experiments show the effectiveness of the proposed method. To obtain the moving objects more accurately, such as to let the edge of tracking region coincide with that of the real object in videos, this dissertation continues to do a series of studies with granular computing.This dissertation generalizes a mapping relationship between semantic mining in images&videos and human cognition from the past algorithms for low-level feature analysis and semantic mining in images and videos. The mapping relationship is:low-level features, knowledge representation and semantic concepts in images&videos to sensation, apperception and presentation in human cognition respectively. So before computers may think like human being, they must acquire sensations from images according to the mapping relationship. It’s a new idea to make computers be able to obtain sensations from a color image through some unsupervised ways. To let the idea come into true, a granule-based partitioning model, based on granular computing(GrC) which is a new way to simulate human thinking to help solve complicated problems in the field of computational intelligence, is proposed for color image processing. First, on the basis of the analyses for rough sets and quotient space, this dissertation constructs a topology information system. And then, this dissertation deems data a hypercube, creates a partition vector set consisting of a basic cross-section vector and several topology cross-section vectors, illustrates the ocular appearances of partitions obtained with the basic cross-section vector and the topology cross-section vectors respectively, defines two new concepts, attribute granules(AtG) and connected granules(CoG), and presents the definitions of the granule-based model. Finally, in order to fulfill the granule-based model, this dissertation designs a single attribute analyser(SAA), defines some theorems and lemmas, describes the processing of extracting all attibute granules in details, discusses limitations, and presents time complexity analysis which shows SAA is high-speed. Experimental results on over300color images show that the proposed analyser is accurate, robust, and able to provide sensations for computers.The sensations obtained above are vague, and in order to make these sensations clear and provide follow-up procedure with a scalable and uniform platform, this dissertation makes a deep study of unsuperivised labeling algorithms for images and designs a granular labeling model suitable for any images even fragments of images. Firstly, this dissertation divided images into concept granules with the granule-based partitioning model according to the selected concept. And then, this dissertation compresses the selected concept granules or all concept granules based on runlength coding row by row and acquires the basic units(referred to as runs) of them. Finally, in order to fulfill the granule-based labeling model, this dissertation designs a labeling algorithm, image granule labeling(IGL), describes the processing of labeling and extracting all connected components(referred to as connected granules) in details, defines the open storage structure which can be used by late stage works, and discusses the expandability of the proposed labeling model to high-dimension. Comparisons and experimental results on binary and color images show that the proposed labeling algorithm is accurate and robust, and quicker than conventional labeling algorithms.Granules used as basic units of the basic platform mentioned above may be connected regions in any shapes or arbitary distributions, which results in a new problem—how to do feature extraction and analysis when basic units are regions. So we makes studies of edge extraction for connected granules. Most of the algorithms for edge extraction are incompetent at edge extraction of arbitrary regions(in any shape or sparse), and without a unified model, the edges obtained by these algorithms are not closely related to the real objects in images. To address these issues, this dissertation construct an edge model related to granules in images based on granular computing. Then the algorithm for arbitrary region edge extraction (referred to as AREE), which is used to fulfill the edge model related to granules in images, is presented. The edge model which is compatible with granular computing consists of four new concepts:topology information system, concept granule, connected granule and edge space. As a way of realizing the edge model, the AREE algorithm introduces some new concepts such as run(a run is a block of contiguous pixels of a concept granule in a row) and interior-point firstly, then presents and proves the theorem used to obtain edges, and finally extracts edges through searching interior-points according to the theorem. The comparative analysis and the experimental results on various types of images show that the proposed algorithm is able to extract edges from arbitrary regions(they may be in any shape or sparse) accurately and quickly.Finally, this dissertation presents an unsupervised rough cognition algorithm based on the basic platform mentioned above for salient object extraction. Firstly, the universe is partitioned with the granular partitioning model according to two concepts respectively, and connected granules are labeled. Then, the following steps are done successively:(1) topology equivalence classes with smaller scale are filtered;(2)the significance of topology equivalence classes is quantized with topology connectivity and topology distribution density;(3)the cut-off position in the significance sequence is found with the improved Fisher’s linear discriminant algorithm, and then candidate regions are obtained by removing non-significant topology equivalence classes;(4)gradual changing pattern is expressed with dimensional scan gradient, which is used to do the local rough segmentation until candidate objects are available, the candidate objects that follow the gradual changing pattern are merged and the significant values of these candidate objects are refreshed;(5)Fisher’s linear discriminant algorithm is run again, and the final salient object is determined according to position weight if more than one candidate object is left. Finally, the executing process of the proposed approach is validated by experiment step by step, and a comparative analysis with three recent methods is conducted, which shows the superiority of the proposed approach in terms of ability to approximate semantic of object and speed.