The particle size and morphology of the rock slag on the TBM conveyor belt are important for the driver’s adjustment of the tunneling parameters. However, the light source at the shooting site produces strong light pollution on the rock image, which greatly affects the image recognition results. This paper proposes an image adaptive spot denoising combined with contour-texture decomposition attention-based method for TBM rock fragmentation classification (AD-CDAN). The proposed method first calculates the global threshold from the grayscale distribution, performs image threshold segmentation to achieve light noise region localization, and then locally fills the texture after image Gaussian smoothing to achieve adaptive denoising. Then, the denoised image is fed into the decomposition attention block, where contour-texture decomposition is carried out for each feature map, and an exactor is presented to optimize the decomposition effect by mapping the decomposition hyperparameters from the input. Meanwhile, the normalized channel attention is computed from the input to output the feature maps with weights. Next, the output results are summed with the features obtained after downsampling the input image, and feedforward processing is performed to obtain the output of a single decomposition attention block. Finally, multiple decomposed attention blocks are stacked and the final extracted image contour features are linearly classified. In addition, this paper proposes a method to add light spot noise to simulate the harsh environment that may occur at the TBM construction site. Experimental validations are conducted on two datasets from the Baolin Tunnel project in Hubei Province and one dataset from Sichuan-Tibet Railway. The results show that the average accuracy of AD-CDAN in the two datasets of Baolin Tunnel without additional noise exceed 93%, and exceed 87% in the two datasets-noised. In the dataset from Sichuan-Tibet Railway, the accuracy of AD-CDAN still exceeds 85%. All results show that the accuracy of AD-CDAN exceeds the comparative models by 2.38%-42.87%, which verifies the effectiveness, superiority, and strong robustness of the proposed AD-CDAN, indicating that the method can be adapted to harsher working environments and provide important support for the safe tunneling of TBM.

中文翻译：

---

基于自适应点去噪和轮廓纹理分解注意力的方法对TBM岩石碎裂进行分类


TBM 输送带上岩渣的粒度和形态对于驾驶员调整隧道掘进参数非常重要。但是，拍摄现场的光源对岩石图像产生了强烈的光污染，极大地影响了图像识别结果。该文提出了一种图像自适应点去噪结合轮廓纹理分解注意力的TBM岩石碎裂分类方法（AD-CDAN）。所提方法首先从灰度分布中计算全局阈值，进行图像阈值分割实现光噪声区域定位，然后在图像高斯平滑后对纹理进行局部填充，实现自适应去噪。然后，将去噪后的图像送入分解注意力块，对每个特征图进行轮廓纹理分解，并提出一个执行器，通过映射来自输入的分解超参数来优化分解效果。同时，从输入计算归一化通道注意力，输出带有权重的特征图。接下来，将输出结果与输入图像下采样后得到的特征相加，进行前馈处理，得到单个分解注意力块的输出。最后，将多个分解的注意力块堆叠起来，对最终提取的图像轮廓特征进行线性分类。此外，本文还提出了一种添加光点噪声的方法，以模拟TBM施工现场可能出现的恶劣环境。在湖北省宝林隧道项目的两个数据集和一个来自川藏铁路的数据集上进行了实验验证。 结果表明，在宝林隧道两个数据集中，AD-CDAN的平均准确率均超过93%，在两个数据集中均超过87%。在川藏铁路的数据集中，AD-CDAN 的准确率仍然超过 85%。结果表明，AD-CDAN的精度比对比模型高出2.38%—42.87%，验证了所提出的AD-CDAN的有效性、优越性和较强的鲁棒性，表明该方法能够适应更恶劣的工作环境，为TBM的安全掘进提供重要支撑。