Abstract: Abstract: To address the issues of missed and false detections caused by the thin, weak continuity of crack structures and the scattered distribution of crack information in images, a crack detection method integrating local texture and global context features was proposed. First, a Complementary Feature Learning Module (CFLM) was designed. In the channel dimension, a uniform partitioning strategy with a 1:1 ratio was used to decompose the features into two groups, which were then fed into the convolutional local perception path and the Transformer global relation modeling path, respectively, to achieve the collaborative expression of local texture and global context throughout the process. Then, a Hierarchical Triple Attention (HTA) mechanism was designed to guide feature information by using channel, spatial and pixel attention to alleviate the inconsistency in semantic response and spatial distribution of feature maps output by dual paths. Finally, an Edge Enhancement Module (EEM) was introduced in the early stage of the decoder. The boundary responses were enhanced by incorporating the Difference of Gaussian and Laplacian of Gaussian operators. Experimental results on four public datasets—Crack500, DeepCrack, ShadowCrack, and AigleRN-TRIMM—show that the proposed method outperforms the other nine comparative methods in terms of F1-score, mean Intersection over Union (mIoU), and thresholded Average Intersection over Union (AIU). The F1-score is improved by 15.26, 5.08, 14.09, and 17.03 percentage points compared to the baseline network U-Net on the four datasets, respectively, and by 0.13, 1.54, 1.43, and 8.08 percentage points compared to the second-best network, respectively. The proposed method effectively improves the accuracy and continuity of crack segmentation results.

Key words: Keywords: crack detection, enhancement of local and global features, convolutional neural network, attention mechanism, feature fusion

摘要： 摘 要: 针对裂缝结构细长连续性较弱、裂缝信息在图像中呈分散分布导致漏检与误检的问题，提出一种融合局部纹理与全局上下文特征的裂缝检测方法。首先，设计了互补特征学习模块（CFLM），在通道维度使用划分比例为1:1均匀划分策略将特征分解为两组，分别送入卷积局部感知路径和Transformer全局关系建模路径，在全过程实现局部纹理与全局上下文的协同表达。其次，设计了多层次三重注意力机制（HTA），利用通道、空间与像素注意力引导特征信息，缓解双路径输出的特征图在语义响应与空间分布上的不一致性。最后，在解码器前期引入边缘增强模块（EEM），该模块引入高斯差分与高斯拉普拉斯算子可增强边界响应。在Crack500、DeepCrack、ShadowCrack和AigleRN-TRIMM四个公共数据集上的实验结果表明，该方法的F1值、平均交并比（mIoU）和阈值平均交并比（AIU）指标均优于其他九个对比方法，F1值较基线网络U-Net在四个数据集中分别提升15.26、5.08、14.09和17.03个百分点，较次优网络分别提升0.13、1.54、1.43和8.08个百分点。所提方法有效提高了裂缝分割结果的准确性与连续性。

关键词: 关键词: 裂缝检测, 局部与全局特征增强, 卷积神经网络, 注意力机制, 特征融合