基于Cascade Mask Region-Convolutional Neural Network-ResNeSt的隧道光面爆破炮孔残痕智能识别方法

Intelligent Recognition Method for Tunnel Smooth Blasting Borehole Residues Based on Cascade Mask Region-Convolutional Neural Network-ResNeSt

为解决现有隧道炮孔残痕识别方法中存在的识别精度不足、鲁棒性较低以及检测速度较慢等问题,提出一种名为Cascade Mask Region-Convolutional Neural Network(Cascade Mask R-CNN)的隧道炮孔残痕识别算法。该算法以Cascade Mask R-CNN实例分割算法为基础,采用先进的Res Ne St网络作为主干网络(Cascade MaskR-CNN-S),增强Cascade Mask R-CNN算法获取特征信息的能力,提升识别的精度;接着采用多尺度训练方式与学习率调整策略对网络进行训练得到智能识别模型,提升识别算法的鲁棒性;最后以平均精度值m AP为测试指标与传统的Cascade Mask R-CNN、Mask R-CNN等算法进行对比试验。研究表明:改进算法的预测框(IoU阈值为0.5)平均精度值(b_m AP(50))与分割(IoU阈值为0.5)平均精度值(s_m AP(50))分别高达0.415、0.350;相较于传统的实例分割算法,改进的算法在隧道炮孔残痕识别精度上有显著提升,隧道爆破残痕长度识别误差仅为8.3%,针对隧道复杂的作业环境具有更好的鲁棒性,抗干扰能力更强。

In order to solve the problems such as insufficient recognition accuracy,low robustness,and slow detec‐tion speed in existing methods for recognizing tunnel borehole residues,an algorithm named Cascade Mask RegionConvolutional Neural Network(Cascade Mask R-CNN)is proposed.This algorithm is based on the Cascade Mask R-CNN instance segmentation algorithm and utilizes the advanced ResNeSt network as its backbone(Cascade Mask R-CNN-S)to enhance the feature extraction capability,thereby improving recognition accuracy.Multi-scale training methods and learning rate adjustment strategies are employed to train the network,resulting in an intelligent recogni‐tion model that enhances the robustness of the recognition algorithm.The model's performance was compared to tra‐ditional algorithms like Cascade Mask R-CNN and Mask R-CNN using mean average precision(mAP)as the evalu‐ation metric.The study shows that the improved algorithm achieves an average precision value of 0.415 for bounding boxes(b_mAP(50))and 0.350 for segmentation(s_mAP(50))at an IoU threshold of 0.5.Compared to traditional in‐stance segmentation algorithms,the improved algorithm significantly enhances the accuracy of tunnel borehole resi‐due recognition,with a length recognition error of only 8.3%.It also demonstrates better robustness and anti-inter‐ference capabilities in the complex working environment of tunnels.