FMANet: Fused mamba attention model with multi-type preprocessing for simulated crack-contaminated complex environments

This paper proposes FMANet (Fusion Mamba Attention Network), a crack segmentation network that integrates Mamba modules and hybrid attention mechanisms, and combines a set of data preprocessing methods that integrate multiple simulated real environment interferences. FMANet enhances segmentation accuracy and anti-interference capabilities with a visual state spatial model and parallel hybrid attention module. The data preparation creates a Chameleon Crack Dataset (CCD) of various cracks using color transformation, Berlin noise, and Gaussian noise/blurring. The experimental findings demonstrate that FMANet obtains 87.38% F1-score and 79.68% mIoU on the CCD test set, which surpasses the other comparison models. The ablation experiment shows the Mamba module’s contribution to the model’s 36.89% improvement in PI value. This work offers an effective way to gather fracture data and automatically segment fractures in complicated settings.