CCMIM: Optimizing concrete defect detection through state-space modeling and dynamic feature fusion

Concrete defect detection is crucial to the safety, reliability, and durability of structures. For CNN models, it is impossible to obtain all information at different scales and complex backgrounds, nor can it capture all contexts globally. Transformer-based models are computationally intensive, making it difficult to generalize to real-time detection tasks. To address these issues, we propose a novel end-to-end concrete crack detection framework: Concrete Crack Mamba-in-Mamba (CCMIM). Specifically, we introduce the Mamba-In-Mamba (MiM) module to capture long-range dependencies and global context to improve the concrete defect detection capability based on hierarchical data flow. In addition, this paper also proposes the Dynamic Dual Fusion (DDF) module, which enhances the robustness and adaptability of the model and achieves smooth multi-scale fusion by dynamically changing the feature representation. To reduce the computational cost and maintain spatial information, we propose the Sparse Pyramid Transformer (SPT) module. This module reduces the computation and improves the inference speed by selecting tokens level by level (from coarse to fine) and sharing attention parameters, but does not sacrifice accuracy. Experimental results show that the CCMIM model outperforms traditional methods as well as YOLO- and Transformer-based models in small crack detection across multiple datasets. Specifically, on the RDD2022, SDNET2018, and CCCD datasets, the accuracy reached 89.2%, 85.2%, and 79.3%, respectively, while the mAP50 reached 88.1%, 87.8%, and 79.2%. In summary, the CCMIM model provides an effective solution for concrete defect detection. The code can be accessed at: https://github.com/lixiaozhen01/CCMIM.