Research On Coal Gangue Image Recognition Method Based On Deep Learning

Coal,as one of China’s traditional energy sources,is closely related to the development of the national economy and people’s lives.In the future,the total demand for coal in China will remain stable.There is a large amount of sulfur element in gangue,and if not handled properly,it will cause serious social harm.Therefore,coal gangue sorting,as the primary issue of efficient coal utilization,has attracted widespread attention and attention from the coal industry.However,due to the high consistency between coal and gangue in terms of appearance,distinguishing coal gangue through the human eye is time-consuming and labor-intensive,and long-term sorting work can also cause significant harm to human health.In addition,traditional convolutional based coal gangue classification networks can only focus on local features of the image,which has certain performance bottlenecks and is difficult to achieve high-level recognition and classification results.In response to the above issues,this article proposes using a Transformer based neural network to achieve classification of coal and gangue.A brand new dataset has been constructed to enable the model to be applicable to different environments,covering various scenes,including ambient lighting,and so on.In order to improve the performance of the object detection network,this article has made improvements on the basis of Cascade-R-CNN and constructed a new model.This model draws on the Cascade-R-CNN network and introduces a neural network based on attention mechanism as a feature extraction network.At the selection level of the backbone network,a large number of ablation experiments were conducted,and ultimately Swin Transformer was used as the feature extraction network.In addition,in order to enhance the performance of the network,a new feature enhancement network "BIFPN+" has been introduced,which draws on the relevant knowledge of residual networks.In the feature enhancement network,a residual structure has been introduced,which improves the performance of the network without increasing the number of parameters.Due to the complexity of coal gangue sorting work,a large amount of data augmentation work was added during the experimental process.The changes in actual environmental light sources can have a huge impact on recognition performance,so data augmentation such as contrast adjustment,brightness adjustment,random cropping,etc.are added to improve model accuracy.Finally,the algorithm was applied to practical industrial environments for testing,and corresponding upper computer software was developed to build an experimental platform.Based on the application scenario of coal gangue classification,a new network framework has been established,and attention mechanism has been introduced as the backbone extraction network in the model.Transformer based networks can better extract global features of images and achieve better recognition results.At the same time,BIFPN+structure has been proposed,which can be plug and play,and is extremely convenient to apply to various network structures and improve network performance.Finally,the overall hardware system was introduced,and a system including the upper computer was produced.After experiments,it was proven that the system structure design was objective and reasonable,with recognition and positioning accuracy.The research results achieved good results in the application of coal gangue image processing,which is in line with the expectations of research experiments.Figure [41] Table [9] Reference [80]...