Research On Coal-rock Image Recognition Method For Complex And Harsh Environment In Coal Mine

There are various types of coal rocks in coal mines,the sudden changes of coal rocks such as faults,ups and downs,and geological structures are complex,dust and water mist are serious,and adverse lighting conditions such as dark light,reflection,and occlusion are prominent.The complex and harsh environment of coal mine leads to the low quality characteristics of underground coal rock images,such as low illumination,weak edge,uneven illumination,low contrast and serious noise interference,which poses a great challenge to the automatic recognition of coal rock.The problem of automatic identification of coal and rock has become a key technology to be broken through in intelligent mining.In view of this,this paper proposes a coal-rock image recognition method based on “enhancement” and“segmentation” for the complex and harsh environment in coal mines.The core technologies include Fuse Transformer and Multi-scale Module Image Enhancement Network(FTM-IEnet)and Fuse Attention Mechanism’s Coal Rock Full Scale Network(FAM-CRFSN).The specific research contents are as follows :(1)In order to realize the enhancement of low-quality coal and rock images caused by complex and harsh environments in coal mines,the FTM-IEnet image enhancement model is constructed based on the color constancy theory and the global feature capture mechanism of Transformer architecture.The model is composed of three network structures : decomposition network,denoising network and brightness enhancement network.The decomposition network is based on Retinex theory,and the overall enhancement of the image is realized by enhancing the illumination component of the low illumination coal rock image.The denoising network uses the NR-Transformer architecture to obtain the dependence between long-distance image information and realize the denoising of the reflection component.The brightness enhancement network uses multi-scale modules to achieve feature capture and fusion of different receptive fields,thereby eliminating local overexposure in the enhanced image.The experimental results show that the FTM-IEnet model effectively enhances the low-illumination coal-rock image,and its PSNR and SSIM scores are 21.288 and 0.783.Respectively,which are much better than conventional image enhancement algorithms such as Retinexnet and RRDnet.(2)The difficulty about automatic recognition of coal and rock in mining face is how to construct a deep learning model that can effectively characterize the semantic features of ’coal ’ and ’ rock ’,and how to achieve the balance between ’ depth ’ stacking and error back propagation.Therefore,this paper takes encoder-decoder as the basic architecture,and takes full-scale connection structure,multi-channel residual attention module with dilated convolution,Res2 Block and multi-dimensional loss function as the core technical elements to construct FAM-CRFSN model,so as to realize pixel-level segmentation of low-quality coal and rock images in coal mine.Among them,the encoder realizes deep semantic feature extraction through convolution pooling module,Res2 module and multi-channel residual attention module with dilated convolution.The full-scale connection architecture is used to enhance the extraction of low-level features.The decoder fuses and analyzes different scale features through bilinear interpolation upsampling,convolution module and full-scale connection structure.The relevant experimental results show that the FAM-CRFSN model achieves accurate segmentation of coal-rock images.When the noise intensity reaches 0.09,85.77% MIOU scores and 92.12% MPA scores are obtained.Its accuracy and generalization performance are much better than Deeplab,PSPNet,Unet,Seg Net and other conventional semantic segmentation models.(3)The above FTM-IEnet model is organically integrated with the FAM-CRFSN model to form a coal-rock image recognition method that first ’ enhances ’ and then ’ segments ’.The relevant experimental results show that the coal-rock image recognition method of“enhancement” and “segmentation” has achieved a score of 98.14% MIOU and 99.11%MPA,which is much better than the score of 60.35% MIOU and 74.87%MPA of the single“segmentation” method,and realizes the pixel-level segmentation of coal-rock images.The research results are helpful to promote the unmanned and intelligent construction of underground mining face.