结合视觉Transformer和CNN的道路裂缝检测方法

Road Crack Detection Method Combining A Visual Transformer and CNN

摘要:

提出了一种结合视觉Transformer和CNN的道路裂缝检测方法。利用CNN来捕获局部的细节信息,同时利用视觉Transformer来捕获全局特征。通过设计的Fusion特征融合模块将两者提取的特征有机地结合在一起,从而解决了单独使用CNN或视觉Transformer方法存在的局限。最终将结果传递至交互式解码器,生成道路裂缝的检测结果。实验结果表明,无论是在公开的数据集上还是在自建的数据集上,相较于单独使用CNN或视觉Transformer的方法,所提出的方法在道路裂缝检测任务中有更好的效果。

Abstract:

This study introduces an integrated approach for road crack detection that harnesses the strengths of both visual transformers and CNN. A CNN is employed to capture fine-grained details, and a visual transformer is fully utilized to capture global characteristics. A feature fusion module is then designed to seamlessly merge the extracted features from both methods, thereby addressing the limitations of using CNN or visual transformer methods separately. Finally, the results are fed into an interactive decoder to produce accurate road crack detection results. Experimental results demonstrate that, whether on a publicly available or self-constructed dataset, the proposed method demonstrates an improvement in performance as compared with using CNN or visual transformer methods separately for road crack detection tasks.