Application Of Image Sorting Of Pb-Zn Ore Based On Adaptive Threshold Semi-supervised Algorithm

With the rapid growth of the demand for mineral products in the national economic development,the development and utilization of mineral resources have reached an unprecedented level.The reserves of high-grade lead and zinc concentrates are decreasing day by day,and the proportion of low-grade refractory lead and zinc ores has increased significantly,resulting in the rising cost of mineral processing and lower economic benefits of enterprises.With the current hot in-depth learning technology,pre separation can be introduced into the process of crushing,grinding and separation of traditional mineral processing technology.It can directly conduct dry separation after the raw ore is crushed.By enriching the concentrate grade,the dilution loss can be effectively reduced,thus significantly improving the comprehensive benefit of the enterprise.However,the deep learning technology of this supervision task relies too much on label samples,which requires a lot of manpower and time cost.This paper collects data based on Dual Energy X-Ray Transmission(DE-XRT)and visible light technology,and performs image preprocessing through improved UNet network segmentation of stones.Finally,this paper proposes a semi supervised depth algorithm with adaptive threshold to solve the problem of label dependency.The algorithm improves the robustness and generalization of the model by randomly enhancing the data and introducing gradient noise,and designs adaptive thresholds for the predicted values of tags and pseudo tags respectively α、β,The former restrains the over fitting of the model,while the latter greatly reduces the generation of noise pseudo labels and improves the accuracy of the model.The backbone network is composed of a residual network and a coordinate attention(CA)module,and the loss is allocated 3:7 through an auxiliary classifier.The experiment shows that the precision can reach 96.37% with 10% labeled samples,which is improved compared with other semi supervised algorithms,and has a performance effect close to 96.96% of the supervised algorithm.The main contents of this paper are as follows:(1)The characteristics and basic principles of ore imaging of dual energy X-ray transmission and industrial camera are introduced.Then,the gray information of the two kinds of data is statistically analyzed.Finally,the advantages and disadvantages of X-ray imaging and industrial camera image are analyzed and compared.(2)A simulation platform is built based on the Pyarch framework,and parallel acceleration is carried out through GPU to facilitate the subsequent training and testing of network segmentation and recognition.(3)The pretreatment process of ore recognition is summarized,in which the improved lightweight UNet network is put forward to achieve real-time segmentation effect.(4)A semi supervised depth algorithm is designed to solve the problem of scarcity of ore annotation data in industrial environment.The training effects of Mix Match,Fix Match,Mean Teacher,Flex Match and other models in the semi supervised field are analyzed and compared through experiments,so as to achieve a sorting effect comparable to supervised learning in the case of a small number of labeled samples and a large number of unlabeled samples.