Fine segmentation method of blast heap block in open pit mine based on point rendering and multi-branch fusion

JIANG Song; RAO Binjian; LU Caiwu; GU Qinghua; RUAN Shunling; YANG Hui

doi:10.13225/j.cnki.jccs.2022.1364

Abstract：In modern mineral exploitation,blasting cost accounts for a large part of the cost of the whole mineral exploi-tation,which makes the control of blasting effect crucial in the whole mineral exploitation process.The control of blas-ting effect is directly related to blasting parameters,which requires the collection of real field data to guide the optimi-zation of blasting parameters.In order to solve the current problems of low accuracy,poor real-time performance and weak generalization performance,a fine segmentation method of blasting pile block under the deep learning framework(Point+S Deeplabv3+)is proposed based on the need of blasting parameter optimization.Firstly,the model introduces a multi-branch separable attention mechanism in the feature extraction part of the backbone network blast pile block degree identification in mining sites,a fine segmentation method for blast pile blocks(Point+S Deeplabv3+)under the deep learning framework to learn the weight features between different channels and combine them,which improves the problem of lack of cross-channel interaction when extracting features.In the decoding stage,the point rendering module is used to gradually splice the low-level semantic features and high-level semantic features corre-sponding to each point by iterating the features of the selected points in different scale feature maps,which solves the problem of losing a large amount of semantic information during sampling,which makes the accuracy of edge and small target segmentation reduced.Finally,the dynamic learning rate adjustment strategy is used to accelerate the con-vergence speed of the model.The experimental results show that the MPA and MIoU of the Point+S Deeplabv3+-based model reach 94.36%and 89.04%,respectively.Comparing with the mainstream semantic segmentation net-works,such as FCN,UNet,PSPnet and Deeplabv3+,the MPA and MIoU of the Point+S Deeplabv3+-based model are improved by 3.04%,4.44%,2.79%,1.52%and 2.95%,4.36%,3.17%,1.88%,with better overall performance,espe-cially for the segmentation of edges and small targets.Therefore,the segmentation model based on the Point+S Deep-labv3+provides a real-time convenient and reliable theoretical basis for the optimal data acquisition of blasting param-eters in the blasting field environment.