Advanced low-light image transformation for accurate nighttime pavement distress detection

Abstract

Pavement distress detection is critical for road safety and infrastructure longevity. Although nighttime inspections offer advantages such as reduced traffic and enhanced operational efficiency, challenges like low visibility and noise hinder their effectiveness. This paper presents IllumiShiftNet, a novel model that transforms low-light images into high-quality, daylight-like representations for pavement distress detection. By employing unpaired image translation techniques, aligned nighttime–daytime datasets are generated for supervised training. The model integrates a lightEnhance generator, multiscale feature discriminators, and distress-focused loss function, ensuring accurate reconstruction of critical pavement details. Experimental results show that IllumiShiftNet achieves a state-of-the-art peak signal-to-noise ratio of 28.5 and a structural similarity index measure of 0.78, enabling detection algorithms trained on daytime data to perform effectively on nighttime imagery. The model demonstrates robust performance across varying illuminance levels, adverse weather conditions, and diverse road types while maintaining real-time processing capabilities. These results establish IllumiShiftNet as a practical solution for nighttime pavement monitoring.