Corrosion segmentation on steel structures via illumination-aware preprocessing and Bayesian U²-Net

Abstract

Steel corrosion poses a major threat to infrastructure safety and longevity, demanding reliable and accurate automated inspection systems. This study presents a semantic segmentation framework that integrates an illumination-adaptive preprocessing pipeline with a Bayesianenhanced U²-Net to address two core challenges in UAV-based corrosion inspection: performance degradation under uneven lighting and lack of predictive uncertainty estimation. Variational Bayesian convolutional layers are embedded in the U²-Net encoder to enhance robustness and regularization in small-sample scenarios without relying on full probabilistic inference. A high-resolution, pixel-level annotated dataset was developed for evaluation. Experimental results across 45 independent training runs show that BU²-Net achieves an F1- score of 75.088%, an IoU of 60.350%, and a recall of 71.537%, while maintaining the lowest standard deviation across all metrics. The results confirm the method’s ability to improve training stability and predictive consistency under visually complex conditions. This practical integration of adaptive preprocessing and lightweight uncertainty modeling supports realworld deployment in multimodal sensing systems for intelligent structural health monitoring.