Research On Coal Gangue Identification And Detection Technology Based On Deep Learning

Mixing gangue with raw coal can lead to low combustion efficiency of coal and generate a large amount of pollutants,so it is necessary to discharge gangue before selecting raw coal.Traditional gangue separation methods not only consume a large amount of material and financial resources,but also have a certain degree of impact on human physical and mental health or the natural environment,and the recognition accuracy is low.With the development of machine vision and artificial intelligence technology,real-time recognition and detection of coal and gangue using deep learning algorithms has become a new research direction.In this thesis,multiple deep learning object detection algorithms are used for analysis and comparison experiments,and the selected algorithm model is improved and optimised for lightness.An efficient realtime lightweight object detection algorithm suitable for low performance detection platforms is proposed.The acquisition and production of image data sets is a key step in the implementation of deep learning object detection algorithms.This thesis builds a coal and gangue image acquisition platform,writes an image collection program to capture images,while adjusting the distance between the industrial camera and the workbench,using a ring light source to ensure the clarity of image acquisition.To simulate the working environment of a coal preparation plant,various types of sample data images were collected,while four different surface types of gangue samples were collected;Then,data augmentation techniques such as spatial transformation,color transformation,and image fusion are used to expand coal and gangue sample data and improve the quality of self-made data sets.Different object detection algorithms,such as Fast R-CNN,SSD,YOLOv3,YOLOv4,YOLOv5,and YOLOX,are analyzed and studied.The four data sets,DATASET-01,DATASET-02,DATASET-03,and DATASET-04,divided in this thesis with different complexity and quantity,are trained using different object detection algorithms.Adjust model parameters,use pre training weights for freeze and unfreeze training on the experimental training platform,and screen the trained models based on the evaluation index F1 score,average accuracy AP,average AP value MAP@0.5,and FPS.The YOLOX network model is eventually found to have good recognition stability for samples from datasets of different complexity,and is able to achieve a MAP@0.5of 99.18% and an FPS of 13.56 with dataset DATASET-04,which is conducive to subsequent lightweight improvements of the model.To improve the detection speed of models,different lightweight models Efficient Net-B0,Shuffle Net V2,Mobile Net series,and Ghost Net are studied.These lightweight network models are combined with YOLOX to replace the feature extraction backbone network of YOLOX model.Testing the performance of network models on a test platform,it is found that the accuracy of the lightweight network models has decreased,with the largest decrease being the YOLOX-Nano algorithm,which is self lightweight,but its detection speed is the fastest;Finally,in order to ensure the real-time detection of the model,Mobile Net V3-YOLOX with a maximum MAP@0.5 of 98.01% is selected as the lightweight network model in this thesisThe MAP@0.5 value of the model has decreased due to the lightweighting improvement of YOLOX,this thesis further attempts to improve and optimize Mobile Net V3-YOLOX to increase the MAP@0.5 value of the lightweight model without significant impact on FPS.Therefore,an attention mechanism with fewer parameters and high flexibility is selected to improve the model.This thesis selects three different attention mechanisms: ECA,CBAM,and CA to replace the SE attention mechanism in Mobile Net V3 or add an attention mechanism to the model to improve the performance of the lightweight model.Among the ways in which lightweight models can be improved using multiple attention mechanisms,it is found that the MB3(CA)-YOLOX network model obtained by replacing the SE in Mobile Net V3 with the CA can increase the MAP@0.5 of the model to 98.38%,which is the best of all improved models,with FPS of 25.8 and meeting the requirements of real-time detection.