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

Area of Science:

• Computer Vision
• Artificial Intelligence
• Civil Engineering

Background:

• Accurate crack segmentation in infrastructure is difficult due to low-contrast, discontinuous defects and environmental noise.
• Existing Transformer models offer global context but have high computational overhead, limiting practical use.

Purpose of the Study:

• To develop an efficient hybrid segmentation architecture for improved crack detection in civil infrastructure.
• To address the limitations of existing models in terms of computational and memory efficiency for real-world applications.

Main Methods:

• A hybrid architecture combining a convolutional encoder (local features) and a lightweight Transformer bottleneck (global context).
• A progressive decoder with multi-level skip-feature fusion to restore spatial resolution.
• Training with a composite Dice-Binary Cross-Entropy objective and a specialized tokenization strategy for fine crack preservation.

Main Results:

• Consistent performance improvements over convolutional, Transformer-based, and hybrid baselines on four public benchmarks.
• Ablation studies confirmed the effectiveness of individual design components.
• Favorable runtime profiling demonstrating low latency and memory usage for real-time deployment.

Conclusions:

• The proposed efficient hybrid model significantly enhances crack segmentation accuracy and efficiency.
• The architecture is suitable for resource-constrained environments and real-time inspection platforms.
• This work contributes a practical solution for automated infrastructure health monitoring.