PEAK-U-Net: A Progressive Edge-Aware Attention KAN U-Net With Sliding Window for Crack Segmentation

Crack detection and quantification in concrete structures remains a critical challenge in structural health monitoring. Traditional manual inspection methods are time-consuming and prone to human error, while existing deep learning approaches struggle with fine-grained edge detection and accurate dimensional quantification. This paper introduces PEAK-U-Net (Progressive Edge-Aware Attention Kolmogorov-Arnold U-Net), a novel architecture addressing these limitations through three key innovations. First, we propose a progressive edge-aware attention mechanism that adaptively focuses on crack boundaries across multiple scales. Second, we present the integration of Kolmogorov-Arnold Networks (KAN) into U-Net architecture, replacing traditional MLP layers with learnable activation functions for superior feature approximation. Third, we develop a multi-scale sliding window framework maintaining sub-millimeter measurement precision based on calibrated pixel resolution. The PEAK-U-Net employs an encoder-decoder structure with enhanced KAN modules in bottleneck and decoder sections, incorporating attention gates and edge-aware convolutions. Our sliding window approach processes images through systematic patch extraction and reconstructs results with maintained spatial accuracy for precise crack width measurements. Extensive experimental validation demonstrates that PEAK-U-Net achieves state-of-the-art performance, particularly excelling in small crack detection with 99.45% accuracy, 81.85% DICE coefficient, and 74.75% Mean IOU on the CrackForest dataset. The method demonstrates superior precision (84.31%) and recall (82.36%) with the lowest loss value (0.21), while successfully classifying crack severity from Hairline (0.07 mm) to Medium (0.20-0.25 mm). PEAK-U-Net represents a transformative contribution to computer vision-based structural health monitoring, offering significant potential to enhance preventive maintenance strategies and provide a robust foundation for next-generation automated inspection systems.