在隧道(巷道)钻爆法施工过程中，智能装药可以取代人工作业，减少装药作业中危险事故的发生。然而，隧道中光线条件差、炮孔目标小和掌子面裂隙等因素会造成智能装药时炮孔的错检和漏检，同时车载计算机有限的算力也是制约炮孔识别大模型使用的难点。MCIW-2深度学习模型，可以解决在隧道掘进作业环境中的高精度炮孔检测和实时性部署问题。模型根据采集到的炮孔图像尺寸特征采取自适应锚框聚类算法优化检测框的长宽比尺寸参数；采用了具有动态非单调聚焦机制的损失函数W_IoU(Wise Intersection over Union)，通过优化边框回归的损失应对低质量炮孔图片的挑战实现了高精度检测；采用了MobileNetv3-Small网络与CBAM(Convolutional Block Attention Module)注意力机制构建了主干网络结构，减少了模型参数保证了检测准确率，满足车载设备的轻量化部署需求。经实验证明，MCIW-2模型在炮孔识别精确率方面达到了96.18%，检测速度达到了59 fps。与基准YOLO(You Only Look Once)系列目标检测模型文件最小的模型相比，所构建的轻量化炮孔智能检测模型减小了75.86%，模型文件仅为2.80 Mb，优于YOLO系列的基准目标检测模型。使用MCIW-2深度学习模型对工作面现场视频进行测试，实现了快速、精确地检测炮孔，测试结果表明，该模型适用于智能装药工程的轻量化部署需求，具有良好的适应性，在综合性能方面具有显著优势。

In the construction process of tunnel (roadway) drilling and blasting, intelligent charging can replace manual operation and reduce the occurrence of dangerous accidents in charging operation. However, some factors such as poor light conditions in the tunnel, small blasthole targets, and cracks in the tunnel face will cause the misdetection and missed detection of blastholes during intelligent charging. At the same time, the limited computing power of the vehicle-mounted computer is also a difficulty that restricts the use of large models for blasthole identification. The MCIW-2 deep learning model can solve the problem of high-precision blasthole detection and real-time deployment in the tunnel excavation environment. According to the size characteristics of the collected blasthole images, the model adopts the adaptive anchor frame clustering algorithm module to optimize the aspect ratio size parameters of the detection frame. The loss function W_IoU (Wise Intersection over Union) with a dynamic non-monotonic focusing mechanism is used to deal with the challenge of low-quality blasthole images for achieving a high-precision detection. The MobileNetv3-Small network and CBAM (Convolutional Block Attention Module) are used to build a backbone network structure, reducing model parameters to ensure detection accuracy and meet the lightweight deployment requirements of vehicle equipment. Experiments have proved that the MCIW-2 model has reached 96.18% accuracy in blasthole recognition, and the detection speed has reached 59 fps. Compared with the benchmark YOLO (You Only Look Once) series target detection model with the smallest file, the lightweight blasthole intelligent detection model constructed is reduced by 75.86%, and the model file is only 2.80 Mb, which is better than the benchmark target detection model of the YOLO series. The MCIW-2 deep learning model is used to test the live video of the working face, and the rapid and accurate detection of blasthole is realized. The test results show that the model is suitable for the lightweight deployment requirements of intelligent charge engineering, has a good adaptability, and some significant advantages in comprehensive performance.