【摘要】 【目标】人工巡检模式效率低下，人力成本高昂，难以覆盖全周期监测需求，且复杂隧道环境下渗漏水区域形态多变，传统图像处理方法难以精准捕捉边界特征等传统隧道巡检过程中存在的问题。针对以上问题，提出一种基于改进DeepLabV3+的隧道渗漏水图像分割方法AF-DeepLabV3+。【方法】首先，使用轻量化的MobileNetV2网络替换原DeeplabV3+模型中的Xception主干网络，以提高分割渗漏水图像的速度，降低参数量。其次，在主干网络生成低级特征之后，设计了一个全局特征增强模块(GFEM),该模块通过改进融合空间和通道注意力的双分支设计，通过通道注意力动态校准关键特征响应，结合空间注意力精准定位渗漏水区域，实现特征表示的互补强化。最后，设计了一个自适应空洞空间金字塔池化(AD-ASPP)模块，该模块通过设置多条空洞率不等的卷积分支和深度可分离的结合对特征进行多尺度的全局上下文的特征融合，可提高模型对渗漏水部分的关注程度。【结果】使用轻量化的MobileNetV2网络降低了模型的参数量，GFEM模块和AD-ASPP模块的结合则增强了对多尺度特征融合的效率，AF-DeepLabV3+模型的渗漏水交并比达到90.24%,模型参数量大小为5.85M,图像处理速度达82.97帧/s。【结论】与当前主流经典图像分割模型相比，AF-DeepLabV3+模型在分割精度和分割速度等方面展现出了优秀的性能。更多还原

【Abstract】 [Objective]Conventional tunnel inspection relies heavily on manual patrols, which are characterized by low efficiency, high labor costs, and limited capability to support full-cycle monitoring. Moreover, the leakage regions exhibit diverse and irregular morphologies in complex tunnel environments, making it difficult for traditional image-processing methods to accurately capture boundary features. To address these limitations in conventional tunnel inspection, the image segmentation method(i.e., AF-DeepLabV3+) based on the improved DeepLabV3+ is proposed for tunnel water leakage detection. [Method]First, the lightweight MobileNetV2 network was adopted to replace the original Xception backbone in DeepLabV3+ model, thereby improving segmentation speed and reducing the number of model parameters. Second, a global feature enhancement module(GFEM) was introduced after the low-level features generated through backbone network. This module employed an improved dual-branch design that integrated spatial attention and channel attention mechanisms. The channel attention dynamically recalibrated critical feature responses, while spatial attention accurately localized leakage regions, enabling complementary enhancement of feature representations. Finally, the adaptive atrous spatial pyramid pooling(AD-ASPP) module was designed, which employed multiple convolution branches with different atrous rates combined with depthwise separable convolutions to perform multi-scale global contextual feature fusion, thereby enhancing the model’s ability to focus on leakage regions. [Result]The adoption of lightweight MobileNetV2 backbone significantly reduces model parameters. The integration of GFEM and AD-ASPP modules further improves the efficiency of multi-scale feature fusion. The proposed AF-DeepLabV3+ model achieves an intersection over union of 90.24% for leakage segmentation, with 5.85M parameters and an image processing speed of 82.97 FPS. [Conclusion]Compared with current mainstream classical image segmentation models, the proposed AF-DeepLabV3+ model demonstrates superior performance in both segmentation accuracy and computational efficiency for tunnel leakage detection.更多还原