Research On Coke Micrograph Segmentation Analysis

The microstructure (including optical texture and pore structure) of coke is closely related to conductivity, thermal expansion, mechanical strength and graphitizability. Analyzing this microstructure accurately and rapidly has important theoretic and practical significance for blast furnace production. Segmentation of coke micrograph is the foundation of automatic classification and recognition for coke optical texture. In this dissertation, according to differences of coke optical texture components in color and texture, texture features is extracted, which is based on co-occurrence matrix. And together with color features, a feature vector is built up for the description of coke optical texture. Then mean shift algorithm is improved and coke micrograph is segmented with it satisfactorily. The main research work in this dissertation is as follows:(1)On extensive relevant literature research, the domestic-international research status of coke optical texture analysis and color image segmentation are summarized. The characteristics and differences among different coke optical textures are analyzed in detail.(2)In the view of visual characteristic of coke image, features of coke micrograph are extracted from color and texture points. Because color features are obtained from even color space and texture features are extracted from a statistics point of view, using gray level co-occurrence matrix method, a feature set that characterizing the coke texture with independent vectors is obtained.(3)A new segmentation algorithm, combining mean shift and edge confidence, is proposed. Firstly, the edge confidence of image pixels is calculated. With the edge confidence the weighting function of mean shift algorithm is computed. The sampling points of feature space are weighted in order to improve the accuracy of detected modes. Secondly, coke micrograph is segmented by iterating the weighted mean shift vector.(4)The coke micrographs are segmented with the proposed method, FCM method, and fuzzy c-means cluster combined with region-growing and hierarchical-clustering methods. By comparing and analysis results of these methods, the validity of our proposed method is proved.(5)An image processing platform is constructed with Visual C++ 6.0, and the algorithms of image pre-processing, feature-extracting and image segmentation are programmed and implemented. It provides an effective and flexible platform for further automatic analysis of coke optical texture.The innovations of the dissertation are as follows: edge confidence is introduced to improve the mean shift algorithm; spatial information and feature information are both used in built the kernel function, which made the relativity of characteristic and space in consideration, and made the segmentation more reasonable.This research work provides a reliable foundation for further analysis and recognition of coke optical texture.