Abstract: Concrete crack detection is essential for maintaining aging infrastructure. However, conventional deep learning methods rely heavily on pixel-level annotations, which are costly and time-consuming to obtain. To overcome this limitation, we proposed a weakly supervised crack extraction framework that eliminates the need for pixel-level labels in our previous work. A ResNet-50 model with a Convolutional Block Attention Module (CBAM) is trained using image-level or patch-level annotations, and Score-CAM is employed to visualize crack regions. In this article, to enhance detection accuracy, a Multi-Scale Patch Fusion (MSPF) module is introduced, which processes subdivided image patches and fuses their outputs, thereby emphasizing local features and improving the detection of fine cracks. Although MSPF enhances crack detection, it can occasionally generate false positives due to background textures. To further address this issue, we introduce a False-Positive Patch Filtering (FPPF) module, an auxiliary CNN that automatically suppresses erroneous detections. Experiments on a benchmark dataset show that the combination of MSPF and FPPF achieves higher recall and IoU than the baseline, enabling robust crack detection without pixel-level annotations. Overall, the proposed framework substantially reduces annotation costs while maintaining high extraction accuracy, providing a scalable and practical solution for automated infrastructure inspection.

Keywords: Crack detection; convolutional neural network; weakly supervised learning; image segmentation; class activation map