An Innovative Crack Detection Algorithm Based on Efficient Feature Fusion and Progressive Transfer Learning

Detecting cracks is a crucial task to ensure the safety of engineering. However, the complexity of the crack background and the highly uneven pixel ratio between the crack and the background significantly limit the effectiveness of detection algorithms. Therefore, an efficient and innovative algorithm for crack detection is proposed, which utilizes a transformer and a multilevel cross-scale weighted feature fusion module, as well as a progressive transfer learning (TL) strategy. Firstly, in terms of existing algorithms, there are limitations in terms of high leakage and false detection rates, as well as a lack of generalization ability, therefore, a F2N-CrackNet model is proposed, which consists of two efficient feature fusion modules and a Nested Multi-level Attention with Atrous Spatial Pyramid Pooling (NMA-ASPP). Secondly, to tackle sparse and unevenly distributed bridge crack data, enhance model detection capabilities, and expedite convergence, a progressive three-stage hybrid TL strategy incorporating across-domain, inter-domain, and inner-domain transferring is proposed, by migrating knowledge from related domains. Finally, to enrich experimental data supporting and validating the proposed model and the transfer learning approach, a new group of experimental data set is offered, comprising hundreds of high-resolution bridge crack images collecting in Shanghai by using a digital camera. Additionally, an efficient human-computer interactive labeling approach combined with morphological processing is devised. Experiments conducted on both open-source and private datasets reveal substantial performance improvements in crack detection achieved by the proposed model and TL strategy. On private data, the model achieves an 81.28% MIoU, exceeding or approaching the performance of state-of-the-art (SOTA) methods, while simultaneously reducing training time by 5–8 seconds per round compared to SOTA. Meanwhile, the performance on multiple sets of open-source data is similarly elevated.