水电能源科学

针对水工混凝土裂缝在复杂背景下形态多变、跨度大且分布极度稀疏等特性，现有基于卷积神经网络的裂缝识别方法存在识别不连续、细节丢失等痛点，本文提出一种基于GCViT-FROS的精细化识别方法。该方法以具备层级化混合注意力机制的Global Context Vision Transformer（GCViT）为编码器，用其“局部-全局”交替感知特性，在精准捕捉边缘细节的同时构建长程依赖。在特征融合阶段，设计改进感受野模块（Modified Receptive Field Block，MRFB）并集成于特征金字塔网络，利用MRFB的膨胀卷积增强空间采样，金字塔结构提升深层信息挖掘能力。在解码阶段设计了空间自适应门控机制（Spatial Adaptive Gate，SA-Gate），融合对象上下文表示模块，利用裂缝区域语义关联及门控机制抑制背景中水渍、麻面等噪声，增强裂缝区域的像素一致性。消融实验结果验证了各核心模块对模型性能的提升作用。在对比实验中，本模型平均交并比达到了86.84%，且本模型识别裂缝的连续性与精度均优于主流模型。在溢洪道该典型水工场景中应用，验证了该方法在实际工程场景下具有良好的泛化能力与应用潜力。

To address the challenges posed by hydraulic concrete cracks under complex backgrounds—namely， high morphological variability， large spans， and extremely sparse distribution—existing convolutional neural network-based crack identification methods suffer from discontinuous recognition and loss of fine details. This paper proposes a refined identification method based on GCViT-FROS. The method employs the Global Context Vision Transformer (GCViT)， which features a hierarchical hybrid attention mechanism， as the encoder. By leveraging its alternating local-global perception capability， the method accurately captures edge details while simultaneously establishing long-range dependencies. In the feature fusion stage， a Modified Receptive Field Block (MRFB) is designed and integrated into the feature pyramid network. The MRFB uses dilated convolutions to enhance spatial sampling， and the pyramid structure improves the ability to mine deep information. In the decoding stage， a Spatial Adaptive Gate (SA-Gate) mechanism is devised， incorporating an object context representation module. By exploiting the semantic correlations of crack regions and the gating mechanism， background noise such as water stains and rough surface textures is suppressed， thereby enhancing pixel-wise consistency within crack regions. Ablation study results validate the effectiveness of each core module in improving model performance. In comparative experiments， the proposed model achieves a mean Intersection over Union (mIoU) of 86.84%， and outperforms mainstream models in both continuity and accuracy of crack identification. Application to a typical hydraulic engineering scenario—a spillway—demonstrates the method’s strong generalization capability and application potential in real-world engineering contexts.