Weak feature crack detection in high-resolution concrete dam surface image with LarkMNet
Jianyuan Li, Xiaochun Lu, Ping Zhang, Qingquan Li
Abstract
Weak feature crack detection is crucial for dam structural health. Existing methods for detecting weak feature cracks in high-resolution images continue to pose challenges. This paper proposes a Perception Large Kernel-based Multi-level Gather-and-Distribute Network (LarkMNet) for weak-feature crack detection in high-resolution images. The construction of a Perception Large-Kernel ConvNet (LarkNet) enables feature extraction and expansion of the receptive field in high-resolution images. The design of the Multi-level Gather-and-Distribute Network (MGDNet) aims to enhance the critical characteristics of weak cracks. Introducing a Shared Detection Head and an Inner-SCYLLA-Intersection over Union further enhances the model’s detection accuracy. Experimental results demonstrate that LarkMNet’s detection accuracy in high-resolution image datasets surpasses that of mainstream detection models, achieving an F1 score and mAP of 89.4% and 93.3%, respectively. Furthermore, the method’s validity has been confirmed through high-resolution image testing and UAV on-site inspections, efficiently and accurately detecting weak feature cracks in high-resolution images.