Research On Ore Image Generation Method Based On Generative Adversarial Network

The application of computer aided technology in ore separation and discarding can effectively improve the classification accuracy of fine tailings.And processing capacity.In recent years,the application of deep learning technology has greatly improved the effect of ore separation and waste disposal,but the implementation of this technology is based on sufficient and balanced training sets,and the application of generative adversance network may become one of the solutions to this problem.In this thesis,through the study of X-ray image and generative adversarial network of lead and zinc ore,a capsule discriminator,generative adversarial network,has achieved the effect of data enhancement to a certain extent.The main work contents are as follows:1.Make ore image data set.In this thesis,X-ray image of lead and zinc mine is selected as the original data,based on the ore data provided by Good Friend Technology,the connected domain labeling method is used to extract the region of interest in the original ore image,and segmentation.Create a data set for subsequent work.2.Analyze the characteristics of ore image data and take into account the training instability of the generative adm network itself,which is prone to mode collapse and generated image blur.Through consulting a large number of domestic and foreign literatures,combining the idea of deep convolutional generative adversarial network and WGAN,and the advantages of capsule network on spatial relative information in small size image classification.A capsule discriminator-generative adversarial network is proposed.The results were analyzed by multiple groups of experiments.It is proved that the model can improve the generation quality and diversity.3.CNN was used to classify ore images for ore tailings,and the discarking rate,concentrate identification accuracy and tailings identification accuracy were used as evaluation indexes to analyze the enhanced effect of generated ore image data.After the optimal model was selected by cross-validation,seven different data sets were designed for comparative tests.CNN classification experiment results show that the generation model proposed in this thesis has good data enhancement performance and can effectively improve the accuracy of the classification model.