Research Of Tungsten Primary Selection Method Based On Multi-feature Fusion

Comprehensive utilization of Tungsten resources is a complicated procedure,and the Tungsten primary selection is the first phase for processing,during which the ore with Tungsten is separated from the rock without Tungsten.Recently the Tungsten primary work is handled by humans,and the workers distinguish the ore with Tungsten from the rock without Tungsten according to the visual information,eg.color,gray scale,texture and brightness.However,the human-based selection is easily influenced by subjective factors,such as criteria difference,visual tiredness,and emotion.And it will greatly increase the cost.Based on the deficiencies with human-based selection,this thesis focused on the basic research on the Tungsten ore primary selection by machine vision.In this thesis,the research works on mineral processing based on machine vision is firstly reviewed.After analyzing the drawbacks in the previous works,regarding to the mineral exploited in Pangu Hill in Ganzhou City,Jiangxi,we have designed a general planning system for Tungsten primary selection based on machine vision.The single feature extracted from Tungsten ore is firstly utilized for primary selection,and then multi-feature fusion will be used for further identification.In this thesis,the main content and contribution is listed as follows:1.Accurate estimation of motion blur parameters.Due to the motion of conveyor belts,the images acquired from cameras are motion-blurred.Traditionally,the motion-blurred direction can be obtained by the light fringe of motion-blurred spectrum image,however it suffered from the influence of cross light fringe and large angle detection error.In order to avoid the cross light fringe influence,we proposed a novel method to obtain the motion-blurred direction using dark fringe in blurred spectrum image,and with the skeletonization and Randon transformation for binary image,the motion-blurred direction is obtained,it will greatly increase the detection precision for motion blur direction and blur length.The average detection error for blur direction is improved from ±2.1? to ±0.13?,and the max detection error is improved from ±4? to ±0.2?.While the average detection error for blur length isimproved from ±2.3 pixels to ±0.2 pixels,and the max detection error is improved from ±2.8 pixels to ±0.3 pixels.With the improvement of estimation accuracy and robustness for motion blur parameters,the recovery quality will also increase.2.Texture extraction.Texture is a very important feature for Tungsten ore identification.Regarding the characteristic of sensitivity to rotation for texture,a novel texture feature extraction method is proposed,the mineral images are firstly handled with OECS-LATP transformation,and then texture can be extracted with Gray-Level Co-occurrence Matrix.With the OECS-LATP transformation,it will greatly enhance the robustness to rotation and the ability to describe the texture.Further more,the texture feature can be extracted with high efficiency and excellent description ability of Gray-Level Co-occurrence Matrix to the irregular texture.3.Design of NP-FSVM classifier on normal plane membership function.With respect to primary classification for mineral images,it has disadvantage of hard classification for the traditional SVM classifier during the training phase,thus we proposed a novel NP-FSVM classifier.During the training phase,the samples were trained in the form of normal plane membership function,which greatly overcome the deficiencies of hard classification and membership function based on cluster centered distance,and dramatically to avoid the influence of abnormal and noisy samplers to membership during the NP-FSVM training phase.From the experiments,it can be observed that the detection accuracy of NP-FSVM is improved by 6.5% and 3.2%,compared to SVM classifier and centered distance membership-based FSVM classifier,respectively.4.Classification based on multi-feature fusion.Due to the deficiencies of low detection accuracy and bad robustness for single feature detection of Tungsten,we proposed a novel detection process.Firstly,NP-FSVM classifier is taken for primary detection for Tungsten color,texture and gray scale.In the following,these three single-features are integrated together via D-S evidence theory,which has the ability to integrate uncertain and incomplete information and advantage of no need of prior probability.Finally,the detection result is obtained by the decision rule of D-S evidence theory.According to the experiments,the detection accuracy is improved by31.5%.Consequently,it overcomes the deficiency of the unreliability,uncertainty,instability in the traditional detection method with single feature,and the detection accuracy and robustness are dramatically improved,thus it is has great impact on the automatic primary selection for tungsten ore.Future work and directions are finally listed after summarizing all the works of this thesis.