Research On Coal Macerals Identification Based On Deep Semantic Segmentation Technology

Coal is the main energy consumed in China and plays a strategic leading role in the national economy.The maceral composition of coal is closely related to its physical and chemical properties,process characteristics,and industrial uses.Its quantitative analysis provides an important basis for comprehensively and scientifically evaluating coal quality and guiding the clean and efficient utilization of coal,and is an important reference index for coal quality and classification.However,traditional methods for identifying coal macerals based on manual analysis have problems such as high labor intensity,strong subjectivity,and low efficiency,which seriously restrict the industrial application of coal macerals.In view of this,this thesis interprets the task of assigning maceral labels to each pixel in coal photomicrograph as a pixel level classification problem,and combines semantic segmentation based on deep learning and multimodal learning theory to study automatic identification method of coal macerals.The main contents are as follows:Aiming at the problem of inconsistent quality of coal photomicrograph data,the histogram equalization algorithm,retinal algorithm,and bilateral filtering algorithm are combined to comprehensively process abnormal samples.The processed coal photomicrograph dataset has improved by 1.53 and 8.68 in information entropy and average gradient,respectively.Then,in view of the limited scale and insufficient diversity of the dataset,various data augmentation methods are combined to expand the dataset,enriching the content and diversity of the dataset.For the problem of coal macerals identification,a macerals identification method based on semantic segmentation and multimodal learning is proposed.Firstly,a variety of classical semantic segmentation methods are used to model coal photomicrograph dataset.The feature extraction network of the Unet model with the best comprehensive performance is optimized and the attention gating module is embedded.A VGG-A-Unet model is proposed.Secondly,aiming at the problem of single modeling data in the existing analysis methods and ignoring the auxiliary role of fluorescent coal photomicrographs,a multimodal fusion method was used to optimize VGG-A-Unet,and an FFS-VAU model was proposed to fully exploit the correlation and complementarity between coal photomicrographs of oil immersed reflective and fluorescent.Finally,aiming at the problem that the feature extraction process of different modal data in FFS-VAU cannot perform information interaction,a cross modal interaction strategy based on skip connection and a redundant information filtering strategy based on efficient channel attention are used to optimize FFS-VAU,and an ESFFS-VAU model is proposed to enhance useful information and weaken useless information while achieving cross modal interaction of information.The experimental results show that the pixel accuracy,mean pixel accuracy,intersection over union,and mean intersection over union of ESFFS-VAU are 94.62%,67.59%,89.78%,and 59.62%,respectively,which are 8.67%,12.68%,12.93%,and15.78% higher than Unet.Considering the current situation that there are few open source automatic identification softwares for coal macerals,a dedicated intelligent analysis software for coal macerals has been developed and designed.The design of the software is based on the Python language and PyQt5 platform,with a user-friendly,concise,and easyto-use orientation,covering functions such as image preprocess,component recognition,and user management.This thesis has 56 figures,8 tables and 98 references.