MultiBlock‐SwinUNet: A Noise‐Robust Framework for Crack Segmentation of Multi‐Class Concrete in CT Images

Considering that industrial CT image segmentation algorithms have low segmentation accuracy when dealing with diverse concrete structures and complex noise environments, especially since multiple sources of interference–such as beam hardening, ring artefacts, metal streak artefacts and partial volume effects in high‐energy CT–further complicate the distinction between cracks and noise, traditional methods struggle to achieve a balance between accuracy and robustness due to the complex interference of noise and artefacts in industrial CT. Therefore, this paper proposes the MultiBlock‐SwinUnet framework to customize and optimize scenes based on functional decoupling and adaptive attention mechanism. Its core is to solve the core contradiction through two problem‐oriented theoretical modules: firstly, the bottleneck module: for the detail‐noise coupling problem under small samples, the three‐level progressive transformer cascade is used to gradually separate crack features and noise based on the multi‐scale feature complementarity theory, and realize the functional decoupling of “feature extraction‐noise suppression‐detail retention”. Secondly, a new node is added in the jump connection stage to introduce resolution‐aware adaptive window attention, and dynamically switch the local/global receptive fields according to the feature map scale to adapt to the multi‐scale characteristics of different concrete structures. And introduce a residual mechanism in the nodes to mitigate the gradient diffusion caused by model depth. The framework alleviates the inherent shortcomings of existing Swin‐UNet in detail‐noise coupling and insufficient multi‐scale modelling through functional decoupling and adaptive mechanisms. The validation of the VoroCrack3D dataset covering multiple concrete types and noise modes shows that the proposed method improves the segmentation accuracy in complex noise scenarios. The average dice coefficient is 90.53%, which is 1.97% higher than the original Swin‐UNet, and the accuracy is improved by 3.45% under high noise conditions (d‐level), which fully verifies the effectiveness of the theoretical design.