Research On Coal Gangue Image Recognition Based On Deep Learning

Coal has important position in the energy structure of our country.In the process of coal mining,a lot of gangue is inevitably mixed,and gangue combustion efficiency is low and pollute the environment,so it is necessary to separate coal and gangue.Accurate identification of coal gangue is a prerequisite for coal and gangue separation,so accurate identification of coal gangue is of great significance.The traditional ray identification method will cause environmental pollution when it is used to identify coal and gangue.Due to the low classification accuracy of coal gangue in image processing technology,the recognition accuracy is not high.The machine learn-based coal gangue recognition method has the problem of low speed due to the need to manually design and extract features.With the continuous development of deep learning technology,the application of deep learning technology in coal identification can quickly and accurately identify coal gangue,thus improving the production efficiency of coal preparation.In the task of coal gangue image recognition,due to the difficulties in coal gangue data collection and processing,the total amount of available coal gangue data sets is small,which easily leads to overfitting of the model.In addition,unbalanced category,variable scale of coal gangue data set and coal gangue easily confused by belt background will affect the final identification effect.Based on the research and analysis of coal gangue recognition at home and abroad,this thesis carries out research in two aspects: data set construction and identification model improvement.The coal gangue data set was constructed to simulate the working condition and the total amount of data was enhanced.Aiming at the shortcomings of existing models,an improved coalgangue recognition model is proposed.The main research work is summarized as follows:A test platform is built to simulate the real coal separation scenario,and coal and gangue sample images are collected and data sets are built.For the problem of imbalance between positive and negative sample categories in coal and gangue images,the data classification sample weights are adjusted,and finally the labeling software is used to complete the label production.The main techniques used for coal gangue recognition are sorted out,and the advantages and disadvantages of the existing research methods are pointed out.For the existing detection models with poor feature extraction ability,large number of parameters and low recognition accuracy,a deep learning based coal gangue image recognition model HPG-YOLOX-S is proposed.first,a lightweight backbone network Ghost Net-S is designed to improve the utilization and feature fusion ability,so that the central region features of the image participate in the classification task and the boundary features participate in the regression task.Then,a hybrid parallel attention module is designed to share detail information across scales,enhance the recognition of object central region information,suppress object edge secondary information,and improve the feature extraction capability of the backbone network.Finally,to improve the localization accuracy of the model bounding box,the SIOU loss function is replaced with the loss function in the original YOLOX-S in the output layer,which can make the model converge quickly.The proposed model is trained,validated and tested on a selfbuilt dataset,and the experiments prove the effectiveness of the improved model.A multi-scale detection model based on Swin-Transformer is proposed for the coal gangue target miss detection problem of the original YOLOX-S detection model in the belt context.First,the shortcomings of the YOLOX-S target detection model are analyzed.The convolutional structure of the original backbone network has translation invariance and is insensitive to the global position of features.For the problem that the global feature information of the coal gangue target is less and inadequately extracted,the Swin-Transformer backbone network with local and global fusion is designed to enhance the identifiability of the target and realize the interaction of local and global information.Then,a multi-scale hybrid dilation convolution is combined with a feature extraction and fusion strategy from multiple scales at all levels to obtain more feature information and improve the detection capability of the target.The Swin-Transformerbased multiscale detection model is obtained by these two improved strategies,which greatly improves the missed detection rate compared with the model in Chapter 3.The experimental results show that the improved model in this chapter greatly improves the leakage detection problem compared to YOLOX-S.This thesis includes 33 charts,10 tables and 78 references.