Research On Deep Learning Algorithm And Key Technology Of Coal Gangue Image Recognition

Coal is a vital strategic resource in China.The resource situation of "rich coal,poor oil,and less gas" makes the coal industry play an essential role in the economic construction of our country.Coal gangue is accompanied by industrial waste in the process of coal production,which has the characteristics of many impurities and low heat.Sensing information from sensors(such as images)to recognize coal and gangue is a fundamental academic problem in coal mining and clean utilization.The coal and gangue separation is one of the essential application directions of coal-gangue recognition.The commonly used separation methods in China are manual separation method,dense medium separation,vibrating screen,and jigging method,and compound dry separation method.These separation methods use more petrophysical and chemical properties to identify coal and gangue,which shall consume many water resources,affect the personal safety of workers,and produce new industrial wastes such as coal slime,which are challenging to conquer.The coal gangue separation methods based on image recognition can merge robot technology and computer vision intelligent algorithm and have the advantages of portable equipment,low power consumption,simple deployment.It has a broad application prospect and great application potential.Intelligent image recognition algorithm has made significant progress in the fields of medical treatment,game,decision-making.Due to the late start in the field of coal gangue image recognition,the related research is not enough.There are still some problems in the image-based coal gangue recognition algorithm,such as the establishment of the visual knowledge graph,the improvement of generalization performance,multi-visual tasks learning.Deep learning methods have solid performance;the computer vision method based on deep learning is a significant part of artificial intelligence.Its computing power and hardware acceleration ability can solve the problem of model generalization and multi-task processing of detection algorithm and can accelerate the research process of image coal-gangue recognition.In this paper,from the artificial intelligence algorithm as the starting point,efforts to solve the coal gangue recognition intelligent algorithm research problems,this paper carries out the corresponding research from the following three aspects.The main results are as follows:(1)the visual knowledge graph of large-scale coal gangue is established with visual cognition as the core.In the existing research on coal gangue identification,large-scale sample collection is delicate,the number of time-consuming data sets is small,and the hidden information such as illumination and sample attitude is not enough,which leads to the lack of generalization ability of the existing algorithms.The progress of intelligent algorithm based on big data driving in the field of coal gangue recognition is significantly limited.At present,the image rendering technology based on the 3D model is the best solution for the expression of image texture and hidden information in the field of image synthesis,so this paper uses semi-automatic modeling technology.In this paper,three-dimensional models of coal and gangue with creamy image texture were constructed,and an automatic image synthesis and labeling algorithm are proposed to quickly establish a large-scale synthetic image data set for the visual cognitive task of coal and gangue.In the experiment,using the real image data set collected in the field,under a large number of coal-gangue image recognition,the effectiveness of the synthetic data set is proved.At the same time,the deep learning algorithm can improve the recognition performance of coal gangue driven by large-scale coal gangue image data set(note:the related work has been submitted to SCI journal,IF=2.707,the detail of research is discussed in the second chapter of this paper).(2)according to the characteristics of coal gangue image,an intelligent recognition algorithm of coal gangue image with high precision and high speed was designed.At present,the running speed of coal-gangue image recognition algorithm is slow,and the recognition method based on depth learning has not been improved.MobileNet lightweight network is one of the most advanced artificial intelligence algorithms,and it has excellent results in image recognition,target detection,and other directions.In this paper,based on the MobileNet convolution neural network algorithm,according to the characteristics of coal gangue recognition,the network structure and loss function are improved,and the large-scale coal gangue image data set in(1)is used.The improved intelligent coal gangue recognition algorithm with high speed and expert recognition accuracy is proposed.This algorithm has a functional phenotype in the experiment,the running speed is only 3ms under the same experimental conditions,and achieves the optimal recognition effect(note:the related work has been submitted to SCI journal,IF=2.707,detailed research is discussed in the third chapter of this paper).(3)the generative adversarial network algorithm was proposed,and the image generator is used to solve the data difference between the synthetic sample and the real sample,to ensure the recognition performance of the algorithm and improve the target detection accuracy of the algorithm.Multi-task learning is the most powerful advantage of the intelligent cognitive algorithm over the traditional image processing algorithm.In this paper,(1)the knowledge graph of synthetic coal gangue image has a variety of visual task markers,and(2)the intelligent coal-gangue image recognition algorithm designed in this paper has a multi-task joint learning mechanism.Through the experiment,the research group found that the detection accuracy of the target detection algorithm based on(2)is not high because of the background difference between the synthetic image and the image sample of the production scene.For improving the quality of the data,the background augmentation generative adversarial network algorithm(BAGAN)and the improved background augmentation generative adversarial with the composite layer(BAGAN-CL)were proposed to enhance the image quality by using the generator part of the algorithm and help the image recognition algorithm(2)in the object detection task to achieve good accuracy(The related work is presented in the EI conference,search number 20184305979063,detailed research is discussed in the fourth chapter of this paper;The related work is published in the SCI journal and discussed in the fifth chapter of this paper).In summary,for the design of the new generation of the artificial algorithm,this paper explores the application of depth learning technology in coal gangue image recognition and solves the research problem of hindering the intelligent coal-gangue image recognition algorithm.It improves the recognition performance and generalization ability of coal gangue recognition algorithm and has important theoretical significance and application value for coal gangue separation technology based on image processing.