Research On The Segmentation Method Of Coal Particles Based On Deep Learning

Coal is an important raw material for production in our country,which occupies a dominant position in energy consumption.In order to effectively remove inorganic minerals related to coal-burning pollution,coal separation is developed to be an important component of clean coal technology.During the coal sorting process,the precise measurement of the particle size of the pulverized coal particles helps to optimize the coal preparation process.With the development of image acquisition equipment and computer technology,the particle size analysis method based on image processing has become an important research direction for its intuitive and accurate analysis results.It can improve the efficiency of coal preparation process evaluation and reduce the workload of traditional manual screening.In this paper,scanning electron microscope and 3D X-ray microtomography are used to study the two-dimensional and three-dimensional particle size distribution of pulverized coal particle,respectively.Since accurate segmentation of coal particle images is the foundation for particle size analysis,this paper proposes methods to segment the particle image based on the characteristics of the two particle images.It is enabled by designing a variety of morphological parameters to characterize the particle size.Through comparison and analysis with the sieving particle size,the most suitable particle size parameters could be obtained to study the particle size distribution of pulverized coal particles.The main research work of this paper is as follows:1.This paper proposes a SEM image segmentation algorithm for pulverized coal particles based on U-Net.It is designed to tackle the problems in processing SEM images of pulverized coal particles where multi-grain particles and particle adhesion often lead to under-segmentation.Firstly,the improved atrous spatial pyramid pooling is used to increase the receptive field of the upper layer features,aiming to effectively segment multi-grain particles.Secondly,the features of multiple scales are merged in the decoding module for the final prediction,which solves the difficult segmentation of smaller particles in the image.Finally,two loss functions are introduced to train the model together to solve the problem of under-segmentation caused by particle adhesion in the SEM image.During training,the particles are assigned greater weights,so that the model can learn more particle information which facilitates effectively segment the adhered particles.The comparative experiment conducted in this work also proves this declaration.At the same time,this paper further analyzes the size of particles based on the segmentation results of the SEM image,via using multiple parameters to characterize the size and shape of the particles.This paper proves the effectiveness of the image-based particle size analysis comparing with the results of the sieving.2.In order to obtain the three-dimensional information of the particles,it is necessary to segment the CT sequence images of coal particles,and then use the three-dimensional reconstruction technology to obtain the three-dimensional structure of the particles.The accuracy of the segmentation of the two-dimensional image sequence determines the accuracy of the three-dimensional reconstruction.This paper designs an end-to-end segmentation model based on Mask R-CNN to obtain individual particle information for CT sequence images.In view of the possible miss segmentation of particles caused by the uneven sizes,this paper adds an attention mechanism to the model to enhance the channel information and semantic information.It is designed to locate small particles more accurately and reduce the problem of missing segmentation.In addition,the segmentation branch is redesigned to integrate the outputs of diverse convolutional layers,which aims to obtain more refined segmentation results when dealing with irregular pulverized coal particle shapes.The experimental results prove the effectiveness of the algorithm in CT image segmentation of coal particles.On the other hand,the usefulness of this method is also confirmed by analyzing the two-dimensional distribution of particles in CT images,and comparing the manually annotated particle sizes with the segmentation prediction of the proposed method.