A Novel CNN–ViT Model with Cascade Upsampling for Efficient Crack Segmentation

Accurate crack segmentation in civil infrastructure imagery remains challenging because of the prevalence of thin, low-contrast, and spatially discontinuous defects that often appear amid textured surfaces, shadows, and acquisition noise. Although Transformer-based models improve global context modeling, many existing solutions incur substantial computational and memory overhead, which limits their use in practical, resource-constrained inspection settings. In this work, we introduce an efficient hybrid segmentation architecture that combines a convolutional encoder for high-fidelity local representation with a lightweight Transformer bottleneck for global dependency modeling, followed by a progressive decoder that restores spatial resolution through multi-level skip-feature fusion. To better accommodate severe foreground sparsity and preserve fine crack structures, the framework is trained with a composite Dice-Binary Cross-Entropy objective and employs a tokenization strategy designed to preserve fine spatial details while enabling efficient global context modeling. We validate the proposed approach on four public benchmarks, demonstrating consistent improvements over representative convolutional, Transformer-based, and hybrid baselines, while ablation studies confirm the contribution of each design component. Finally, runtime profiling shows favorable latency and memory characteristics, supporting real-time or near real-time deployment on embedded and edge inspection platforms.