Studies On Image Semantic Model And Image Semantic Object Segmentation

This dissertation deeply studies image semantic theory and its some applications. Image semantics is a new field in image understanding. Its content consists of low-level feature and high-level feature, and how to map low-level feature to high-level semantic. Now, the theory of the image semantics has few been reported in the literatures. This dissertation mainly studies semantic information of image structure, the problems of how the observers acquire information from images, how to define this kind of information, propose image structure model, color image segmentation and extract semantic object and region based on color feature by clustering algorithms.The main works are shown as the following: 1 We derive strictly an expression of image structure information, and define and extract a structure parameter set and therefore form a structure space T. We map T onto structural information space I_T by a tranformation function set φbased on psychometric functions. denotes the structural information of the given image obtained by a human observer who has understood the structures of an image. We also express a stronger response of a person to a special stimulation I_T by a function δ. We point out that only after analyzing the structure of a given image, can we obtain the structural information,when an observer knows more structural information, and he can make more correct decision on the semantics of that image structure. 2 We first introduce the human psychometric function,secondly,discuss psychological response for image color and shape of images. In the end, we construct a simplified receiver model approximately to human perceptual responses. We try to connect the subjective perpetual with signal processing and communication theory. We finally point out the realization of this model. 3 By employing the evolution model of one-dimension signals, an evolution model of images is constructed in a generalized space, and the relationships between the parameters of the proposed model and the physical variables of the image are given, on the basis of which a structural information model of the image is proposed, H(tran) being defined as the structure entropy of the image. It is indicated that images should evolve to achieve the highest structure entropy. Moreover, a hierarchical structure model is established,all terminal nodes of which form a semantic string. The semantic code of the image is finally defined. The information quantity and entropy of this code are respectively considered as the semantic information and the semantic entropy. 4 We extract semantic object based on color feature using clustering analysis. In this paper we firstly translate RGB space into space. Secondly, we use the nearest neighbor rule algorithm solve clustering problem. At the end, we extract semantic object in terms of color information. Experimental results show that the color clustering give better results in increases in cluster compactness. Lab5. We present a novel image segmentation algorithm, which combines edge and region-merged based techniques. First, an edge-preserving statistical noise reduction approach is used as a preprocessing stage in order to compute an accurate estimate of an image gradient. Second, we segment image into primitive regions by applying watershed transformation on the image gradient magnitude. At the end, we merge neighboring region into homologous region by morphological erosion and dilation. 6. We analyze objectivity and subjectivity model of image. We propose semantic analysis model combining objectivity semantic information with subjectivity perception. 7. A semantic-based image retrieval that can query large image databases efficiently by color, shape and texture. Color histogram is an important technique for image indexing and retrieving. Histogram is simple and invariant for translation and rotation of the color. A new semantic-based image classification method is proposed, in which RGB histograms and the dominant triple are extracted from quantized HSV joint histogram in the local image region. Experiments indicate that this method is effective in image's classification.