Abstract: Fine-grained cracks, as an early stage of road surface crack formation, can be detected and repaired to eliminate safety hazards and reduce maintenance costs in a timely manner. Fine-grained cracks not only have complex topological structures, but also have geometric characteristics of small width and variable scale. In complex road backgrounds, existing methods are prone to missed detections and have low accuracy in perceiving their width. In response to this, this paper proposes a fine-grained crack detection method for road surfaces based on dual branch selective fusion and multi-scale semantic enhancement. An enhanced self-attention mechanism and a dual branch parallel backbone network of CNN (convolutional neural network) are designed to simultaneously extract features from both local and global perspectives, enriching feature representations layer by layer. A redundancy reduction and feature selective fusion (RSF) module is proposed to achieve the learning and interaction of dual branch global and local information, enhancing the expressive power of features. A multi-scale semantic enhancement fusion strategy is adopted to enhance the model’s perception ability of fine-grained crack features through cross scale information transmission and fusion. To validate the effectiveness and reliability of the proposed method, training and testing are conducted on the CrackTree260 public dataset, and the model’s generalization performance is evaluated on the CRKWH100 dataset. The experiment shows that the proposed method achieves ODS values of 0.909 and 0.918 on two datasets, respectively, which is superior to other advanced crack detection methods.

Key words: fine-grained crack detection, self-attention mechanism, convolutional neural network (CNN), multi-scale feature fusion, semantic enhancement

摘要： 细粒度裂缝作为路面裂缝形成早期阶段，对其进行检测和修复可以及早消除安全隐患，降低维护成本。细粒度裂缝除拓扑结构复杂外，还具有宽度微小、尺度多变的几何特征，在复杂路面背景下，现有方法容易出现漏检且对裂缝宽度感知精度不高的问题。针对此，提出了一种基于双分支选择性融合与多尺度语义增强的路面细粒度裂缝检测方法。设计了增强自注意力机制和卷积神经网络（convolutional neural network，CNN）的双分支并行主干网络，从局部和全局同时进行特征提取，逐层丰富特征表示；提出了冗余减少和选择性特征融合（redundancy reduction and feature selective fusion，RSF）模块，实现双分支全局和局部信息的学习和交互，增强特征的表达能力；采用了多尺度语义增强融合策略，通过跨尺度的信息传递和融合，提升模型对细粒度裂缝特征的感知能力。为了验证该方法的有效性和可靠性，在CrackTree260公共数据集上进行了训练和测试，并在CRKWH100数据集上评估模型的泛化性能。实验表明，所提出的方法在两个数据集上分别达到了0.909和0.918的ODS值，优于其他先进的裂缝检测方法。

关键词: 细粒度裂缝检测, 自注意力机制, 卷积神经网络（CNN）, 多尺度特征融合, 语义增强