| As a support for China's economic development,mineral resources are inseparable from environmental protection,energy,and process optimization.According to statistics,more than 90%of China's energy,over 80%of raw materials,and surpassed 70%of production materials are supplied by mineral resources,providing a basic guarantee for China's production and life.In order to implement the "green mine" concept,it is necessary to strictly follow the requirements of mine construction,scientific and rational utilization to ensure the sustainable development of the mining area.At present,the cost for mining areas is higher than in the past,and the quality of minerals is not as good as before,but the market requirements are gradually increasing.Therefore,relevant technical personnel are required to provide professional guidance and classify minerals as needed.The accuracy of the detection and the design of scientific and orderly programs can improve the effective utilization of mineral resources,escort for subsequent product sales and value-added,and ensure economic benefits.Image processing technology provides an effective and accurate means for mineral detection,providing timely and effective methods for mineral resources in China,reducing waste of resources and environmental pollution,and accurately guiding the development of nonferrous metals.However,there are still several problems in the analysis of ore images:1)Strong noise problem in the ore image.Because the ore texture is complex and chaotic,the differentiation of ore objects is low.The internal impurities and noise are wide distributed and high density.The noises have various types,affecting the extraction features and the determination of the particle size category;2)Adaptability of particle size classification is low.The ore is crushed step by step by the action of the crusher,resulting in large differences in ore characteristics,and the traditional image processing methods are difficult to detect.3)Research on the accuracy of ore particle size distribution in complex scenes.The ore particle sizes are different,and the ore block is large and covered by mud,etc.The accuracy of particle size distribution is affected by the small size ores covered by dust and sand which make the ore features hard to extract.This topic provides a new perspective for the distribution and identification of ore particle size,and has important significance in the fields of mineral processing,metallurgy,chemical industry,etc.It provides guarantee for intelligent and automated particle size distribution and identification.As a self-selected topic,this thesis focuses on the problems of ore images,combines the external collection environment,focuses on ore segmentation,spatial domain and frequency domain feature extraction,and classifies the images under different broken level,and establishes a combination vector based on entropy,to complete ore identification in complex scenes.The main contents of the thesis include:1)Spatial domain feature extraction model based on ore image multivariant multiscale explores the method of non-segmentation.The ore image features are extracted at spatial domain.2)Frequency domain feature extraction model based on two-dimensional empirical mode decomposition of ore image.The ore image feature extraction at the spatial frequency domain.3)Dynamic optimization ore segmentation algorithm based on histogram accumulation moment,which accurately separates ore targets from the background pixels.4)Complex ore image recognition model based on entropy combination vector is used for ore image recognition.The innovations of this paper have the following aspects:(1)In order to solve the ore image objects detection problem of illumination and overlapping particles,a multiscale feature extraction model based on ore image spatial domain is proposed.Image color and gradient information are extracted at multiple scales,and composite delay vector is constructed on line space.The vectors are constructed by spatial features at multi-scale.The Chebyshev distance is used to find the maximum spacing between two vectors.The distance is used as representations of scale features and used for ore identification feature.By reflecting multiscale particle size information,it overcomes the problem of ore segmentation.(2)The spatial domain feature is difficult to comprehensively and effectively represent the ore image characteristics.A frequency domain feature extraction model based on two-dimensional empirical mode decomposition of ore images is proposed.Two-dimensional empirical mode decomposition extracts local frequencies at different broken level ore images,reflecting the edge,texture and other information of image objects.The calculation of the instantaneous frequency of each modal image quantifies the frequency domain information at different scales,comprehensively reflecting the boundary,texture and roughness information of ore surface and constituting effective features for ore particle size recognition.(3)Aiming at the problem of existing threshold segmentation,an ore segmentation algorithm based on histogram accumulation moment is proposed.The zero order cumulative moment of the histogram reflects the gray level probability and the first order cumulative moment represents the mean value.Compared to the bi-neighbour OTSU,the algorithm reduces the internal noise of the target.In the entire image,the algorithm is used for the segmentation of ore image.By calculating the 0th moment representing the gray level probability and the first order cumulative moment representing the mean near the target and the background part,the method is based on OTSU,instead of calculating the entire gray level distribution,reducing the error caused by noise.By choosing the optimal threshold to avoid falling into local optimum,it is not only suitable for the histogram image of bimodal distribution,but also effectively segments the single or near single peak ore image to realize ore segmentation quickly,effectively and accurately in complex scenes.(4)An ore particle size distribution model based on entropy combination vector is proposed for ore image recognition in complex scenes.From the segmented ore object,the morphological characteristics of objects are extracted,and the particle size information in ore image is calculated.The multiscale features extracted from the spatial and frequency domain are respectively measured by multivariate multiscale entropy,and the maps of entropy values range and particle size distribution on multiple scales are established respectively.The entropy vectors of the spatial and frequency domains are combined to map the relationship between the entropy values and the mixed ore image.The ore image recognition in complex scenes is completed.In this paper,the multiscale ore particle size distribution model has been studied and explored,but it has the following shortcomings.Further research and improvement are needed:l)Based on the limited image spatial domain feature,more features are needed to be extracted to improve the detection accuracy.2)The frequency domain level reflects the ore boundary,texture and roughness information,and subsequent work will consider adding directions and other features for further analysis.3)At present,this paper only deals with the stone with equal mass ratio,and ore images at more complex scenes need to be detected and identified. |
PDF Full Text Request