Research On Content-Based Image Similarity Measure Techniques And Its Application In Hydraulic Engineering

As very large collections of images are becoming common, there is a growing interest in image database that can be quired based on image content. Content-Based Image Retrieval (CBIR) has become an important research issue aiming at providing effective means for image retrieval on large image databases. In the dissertation, content-based image retrieval methods are studied based on information theory and fuzzy set theory, and several techniques, information entropy-based image retrieval, similarity measurement of image entropy, fuzzy index, dynamic relevant feedback are presented. In the end, the research results are applied to design and implement the content-based engineering image retrieval system, and some key techniques are presented for inquiring about flood prevention image information.The main contributions of this research include:1. According to principles and human perception of color, a HSL (Hue Saturation Light) color space, which is perceptively consistent with human vision, is selected and divided. The similarity between two colors is re-defined in order to accord with human discernment for colors. The method of extending color histogram is presented to match images with different colors.2. Defining the entropy space of image and entropy difference, the concept of information entropy is applied to image retrieval. Some mathematical properties of entropy are studied, and similarity measurement of image entropy and corresponding algorithm is presented.These techniques can reduce the dimensionality of histogram space from n to l(n>l), increase the image retrieval efficiency, and improve the capability of image retrieval system.3. Based information entropy, multi-precision similarity matching and multi-dimension hash index are proposed. The image is divided into a number of sub blocks, each with its associated color histogram and image entropy. Theoretics and experimental results show that the techniques can help to make similarity matching more precise.4. A fuzzy image data model and a concept of fuzzy space are proposed, in which model visual feature, spatial feature and semantic feature are used for super feature in order to utilize advantage of traditional relation database as well as characteristics of image data and fuzzy retrieval. Based fuzzy space, a method of similarity measurement of image is presented to support fuzzy features-based image retrieval and satisfy user's query requirement for image. In the thesis, a semantic template and the mechanism of dynamic relevant feedback are defined so that it can express user's query semantic and improve retrieval precision and useable capability for image retrieval.5. A mixed data model called relation-level-object, which is used to image retrieval system in flood prevention, is presented in order to satisfy special access requirement for image data and spatial data. With the research results of this dissertation, a prototype system, a content-based engineering image retrieval system, is designed and implemented.