Multi-target Recognition Of Microscopic Image Of Blast Furnace Dust

Blast furnace dust is an important by-product of blast furnace smelting,which contains large amounts of combustible carbonaceous substances and abundant metal resources.The content of carbonaceous substances can reflect the rationality of blast furnace feeding structure and the stability of production parameters,and metal and metal oxides can be recycled utilized.Therefore,quantitative analysis of blast furnace dust composition is of practical significance to guide the smooth and low-consumption production of blast furnace,as well as the efficient and clean utilization of blast furnac dust.Based on the analysis of the characteristics of target components in blast furnac dust microscopic images,this dissertation employs the approach of combining segmentation and classification to extract carbonaceous substances and metal oxides for recognition.In view of their different sizes and the difficulty of recognition,different recognition schemes are formulated,and the multi-target recognition of blast furnace dust microscopic images is realized by fusing the recognition results.The main contents of this dissertation are as follows:(1)After consulting relevant literature and spot investigation,the research status of blast furnace dust processing,image segmentation and classification methods in target recognition are summarized,and the research background and significance of this dissertation are expounded.(2)The characteristics of target components in blast furnace dust microscopic images are analyzed,and the holistic approach of target recognition is determined according to the differences between and within the target components.(3)In the light of the shortcomings of MeanShift clustering results,MeanShift clustering is improved by combining SLIC superpixel algorithm and region merging algorithm,and an image segmentation method for blast furnace dust microscopic image,which is based on the improved MeanShift clustering,is proposed,the effectiveness of the improved method is verified by comparing and analyzing the segmentation results with that of other methods.(4)In view of the irregular shape and limited feature extraction of segmented regions,approach of feature extraction such as texture of irregular regions under limited conditions is studied,a total of 33-dimensional features about color,texture and edge of irregular segmented regions including carbonaceous substances and background region are extracted for visual analysis,the feasibility and accuracy of the feature extraction method for irregular regions are verified by experiments.(5)Based on the non-linear discriminant ability and the suitability for the small sample classification of support vector machine(SVM),clalssifier based on SVM is designed,and segmented regions are classified by classification schemes including the combination,dimension reduction of features and “the binary tree”,and the optimal classification scheme is selected to identify carbonaceous substances.(6)Aim at the characteristic of intensity of metal oxides and the shortcomings of recognition by threshold segmentation,an recognition scheme of “fixing contour + determining position + eliminating interference” for metal oxides is formulated,and the scheme is realized based on super-pixel segmentation,rules classification and neighborhood differences,and the effectiveness of the improved scheme are verified by the recognition results.The special and innovation of the dissertation mainly lie in: Using image segmentation and classification methods to analyze blast furnace dust from micro-perspective;Putting forward an image segmentation method based on improved MeanShift clustering to implement a fast and effective segmentation of carbonaceous substances;Giving a solution to extract the relevant features of irregular regions,and carrying out classification experiments from different classification perspectives combined with multiple features;Putting forward a metal oxides recognition scheme,which achieves effective target recognition and can be used as a reference for recognizing other targets with high brightness and small size.