煤矿巷道断面裂隙检测可表征巷道地质情况，指导掘进与支护工程。传统的巷道断面裂隙图像处理方法容易受到煤矿井下复杂环境的影响，导致裂隙检测效率较低。

提出一种基于DC-UNet模型的巷道断面裂隙检测方法，可部署在移动端与嵌入式设备，实时有效地检测巷道断面裂隙情况，提高裂隙检测效率。首先，建立了巷道断面裂隙图像数据集，使用图像数据增强技术增加原始图像数据集，并对数据集图像进行标注。随后，基于DC-UNet网络构建了裂隙检测与裂隙参数计算框架，通过改进深度可分离卷积(depthwise separable convolution，DwConv)与引入双重注意力模块(convolutional block attention module，CBAM)，提高了模型的轻量化与检测精度。DwConv替代传统卷积操作，降低了模型的参数量和计算量；CBAM模块通过串联通道注意力与空间注意力，提高了模型在低照度多粉尘复杂井下环境中的裂隙语义信息。最后，将所提算法与现有图像处理算法进行了对比，并将模型导入到移动端设备中进行验证。

结果表明：改进模型的裂隙检测精度为92%，相比于传统图像处理方法，检测精度得到了提高；改进模型部署在移动端设备，数据量为7.52 MB，相比于原有模型减少了68.9%，降低了模型的计算量；占用空间为19.43 MB，相比于原有模型降低了80.2%，提高了模型的轻量化水平；检测时间为0.075 s，满足现场检测实时性的要求，提高了检测速度。提出的改进算法可以满足煤矿井下掘进断面裂隙检测要求，为煤矿巷道裂隙检测与巷道掘进工程奠定了基础。

Detecting fractures along roadway sections of coal mines allows for the characterization of the geologic conditions of roadways, thus providing guidance for roadway tunneling and support engineering. However, conventional image processing methods for fractures along roadway sections are susceptible to complex environments in underground coal mines, exhibiting a low fracture detection efficiency.

This study proposed a novel image processing method for detecting fractures along roadway sections based on a DC-UNet model. This method can be deployed on a mobile terminal and an embedded device, thereby allowing for real-time and effective detection of fractures along roadway sections and enhancing detection efficiency. First, an image dataset of fractures along roadway sections was established. The dataset was then enhanced using image data augmentation techniques, and the images in it were labeled. Subsequently, a framework for the detection and parameter computation of fractures was established based on the DC-UNet architecture. The improved DC-UNet model featured significantly elevated lightweight nature and detection accuracy by improving depthwise separable convolution (DwConv) and incorporating the convolutional block attention module (CBAM)—a dual attention module. Specifically, DwConv was employed to replace conventional convolution operations, reducing the model's parameters and computational load. The CBAM, integrating channel attention with spatial attention, enhanced the model's ability to capture semantic information related to fractures in low-light, dusty, and complex underground environments. Finally, the proposed method was compared with existing image processing algorithms, and the results were validated by importing the DC-UNet model into mobile terminal devices.

The results indicate that the improved model achieves a fracture detection accuracy of 92%, higher than that of conventional image processing methods. The improved model, after being deployed on mobile devices, shows a data size of 7.52 MB, which is reduced by 68.9% compared to the original model, suggesting a decreased computational load. This model occupies a space of 19.43 MB, representing a reduction of 80.2% compared to the original model, suggesting an enhanced lightweight nature. This model exhibits a detection duration of 0.075 s, satisfying the requirements for field real-time inspection and accelerating the detection speed. Overall, the proposed improved method can be applied to detect fractures along the roadway sections in underground coal mines, laying a foundation for the fracture detection and tunneling engineering of roadways in coal mines.