CrackNex: a Few-shot Low-light Crack Segmentation Model Based on Retinex Theory for UAV Inspections

Routine visual inspections of concrete structures are imperative for\nupholding the safety and integrity of critical infrastructure. Such visual\ninspections sometimes happen under low-light conditions, e.g., checking for\nbridge health. Crack segmentation under such conditions is challenging due to\nthe poor contrast between cracks and their surroundings. However, most deep\nlearning methods are designed for well-illuminated crack images and hence their\nperformance drops dramatically in low-light scenes. In addition, conventional\napproaches require many annotated low-light crack images which is\ntime-consuming. In this paper, we address these challenges by proposing\nCrackNex, a framework that utilizes reflectance information based on Retinex\nTheory to help the model learn a unified illumination-invariant representation.\nFurthermore, we utilize few-shot segmentation to solve the inefficient training\ndata problem. In CrackNex, both a support prototype and a reflectance prototype\nare extracted from the support set. Then, a prototype fusion module is designed\nto integrate the features from both prototypes. CrackNex outperforms the SOTA\nmethods on multiple datasets. Additionally, we present the first benchmark\ndataset, LCSD, for low-light crack segmentation. LCSD consists of 102\nwell-illuminated crack images and 41 low-light crack images. The dataset and\ncode are available at https://github.com/zy1296/CrackNex.\n