基于点渲染的多分支融合露天矿爆堆块体精细分割方法

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

现代矿产开采中,爆破成本控制占据整个矿产开采的成本控制的很大一部分,这使得爆破效果的控制在整个矿产开采流程中至关重要,而爆破效果的控制与爆破参数直接关联,这需要采集现场真实的数据来指导爆破参数的优化。为解决当前对于矿区现场爆堆块度识别时存在的精度低、实时性差、泛化性能弱等问题,基于爆破参数优化的需要,提出了深度学习框架下的爆堆块体精细分割方法(Point+S Deeplabv3+)。首先模型在骨干网络特征提取部分引入多分支可分离注意力机制,学习不同通道间的权重特征并融合,改善了提取特征时跨通道交互缺乏的问题;在解码阶段,运用点渲染模块,通过迭代选取点在不同尺度特征图的特征,逐步对每个点对应的低级语义特征和高级语义特征进行拼接,解决了上采样时大量丢失语义信息、边缘及小目标分割精度降低的问题;最后使用动态学习率调整策略,加快模型的收敛速度。实验结果表明,基于Point+S Deeplabv3+模型的MPA和MIoU分别达到了94.36%和89.04%,对比主流的语义分割网络,如FCN、UNet、PSPnet和Deeplabv3+相比,基于Point+S Deeplabv3+的模型MPA和MIoU分别提升了3.04%、4.44%、2.79%、1.52%和2.95%、4.36%、3.17%、1.88%,具有更好的综合性能,特别对于边缘和小目标的分割效果有明显改进。因此,基于Point+S Deeplabv3+的分割模型为在爆破现场环境下的爆破参数优化数据采集提供了实时便利、可靠的理论依据。

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.