Research On Microscopic Mineral Phase Feature Extraction Method Based On Deep Learning

Mineragraphy is of great significance to the exploitation and utilization of mineral resources such as mining,beneficiation,and the chemical industry.The general steps of traditional mineralogy research are: judging the type of ore,analyzing the mineralogy under the microscope,and arranging the data.Given the low efficiency and high experience requirements of traditional mineralogy research,exploring the automation of ore detection and ore microscopic image segmentation.There are 1800 data sets of 8 types of ore,The target detection network is built with the lightweight model Shuffle Net V2 as the backbone,The dynamic positive and negative sample allocation strategy is used to make the positive samples allocated in the training phase more accurate.For the problem of class imbalance,Using progressive equalization sampling to improve the accuracy of few-sample data,and the distillation learning strategy is used to improve the generalization of the model.The final mean average precision reached50.5%.To optimize the generalization of the classification model and prevent the classification model from classifying according to the unique gangue in the mineral facies.The ore facies is divided into gangue area and main mineral area by threshold and area of interest segmentation strategy,and the data set is added for training,so as to guide the model to pay more attention to the main mineral.Increased attention.Aiming at the difficulty of manual labeling caused by the complex mineral phase texture features and dense particles under the microscope,taking the magnetite with the residual structure as the research object,Improve U-Net by fusing multiple dimension feature maps,and the improved model FLOPs is only 4.72 G,and the model obtained by training with 4 samples can segment the outline of the quartz area.After median filtering,the m AP reaches 85.6% on the validation set containing 20 images.The edge detection operator was used to segment the magnetite grain region metasomatized with quartz,and the magnetite grain region metasomatized with galena was obtained by threshold segmentation.Figure 58;Table 6;Reference 67...