ULiteCrackNet: An ultra-lightweight crack segmentation model balancing accuracy and efficiency with stable diffusion–enhanced transfer learning

Training lightweight segmentation models for detecting cracks from images involves a trade-off between accuracy and model size, especially when limited training data are available. This paper presents an ultra-lightweight segmentation method – ULiteCrackNet, which achieves high accuracy while maintaining minimal computational cost. ULiteCrackNet advances the field of lightweight networks by introducing a backbone that synthesizes key modules: an Edge-Aware Attention Module, Mobile Inverted Bottleneck Convolutions, and an Ultra-Lightweight Subspace Attention mechanism. This integrated design achieves superior performance compared to current lightweight methods, achieving an IoU of 88.8% and an F1 score of 94.1%, requiring only 1.46 giga floating point operations per second, and an inference time of 3.3 ms per image. To further enhance the performance of the lightweight model and address issues of limited data and bias, a two-stage transfer learning strategy is employed, combining weakly supervised pre-training with fully supervised fine-tuning. The novel aspect of this strategy is the use of a Stable Diffusion inpainting model to generate synthetic images. These synthetic images then serve as coarse labels for pre-training, effectively mitigating data scarcity. Results show that IoU improves from 88.8% to 89.8%, and F1 score increases from 94.1% to 95.0%. The ablation studies further validate that the developed lightweight model, when integrated with coarse-to-fine transfer learning, demonstrates a clear capacity to mitigate the real-world engineering challenges of limited datasets and inherent data bias in structural health monitoring.