| The chute is the key channel for coal transportation in the coal preparation plant.The blockage of the chute is a common fault in the production process of the coal preparation plant.Existing detection technologies are all based on contact detection methods,and there are still certain problems.Image processing can be used for anti clogging detection,but there are few related studies.In order to improve the efficiency of the chute anti clogging detection,this thesis explores the task of denoising coal chute image and the task of segmenting coal flow contour,the main research contents are as follows:(1)This thesis proposes a denoising algorithm NL-DIP for deep image prior based on non-local regularization.NL-DIP integrates the non-local self-similarity of the image on the basis of the deep image prior,so as to make full use of the prior knowledge of the image to denoise.The experimental results show that the evaluation indexes PSNR and SSIM values of NL-DIP on the standard test set set5 have been improved,effectively verifying the denoising performance of NL-DIP.When suppressing the blind noise of the coal chute image,NL-DIP can not only achieve effective denoising,but also better retain the texture information such as the edge details of the coal flow contour.(2)This thesis proposes a feature integrated semantic segmentation network FIDeeplabv3.The network uses a Multi Scale Pooling Integration Module to connect the output of each convolutional layer in the encoding stage to the largest pooling layer of different scales to realize the translation invariance of the feature mapping of each submodule.And a Decoding Residual Module is used to realize the fusion of the shallow features and deep features of each submodule in the decoder.The experimental results show that FI-Deeplabv3’s evaluation index Io U and PA values on fuzzy coal image data are higher than those of excellent networks such as Deeplabv3,UNet,PSPNet,Seg Net.The edges of coal contours obtained by image segmentation are more accurate,effectively reducing the problems of over-segmentation and undersegmentation.The thesis has 38 pictures,7 tables,and 84 references. |
PDF Full Text Request