A Bio-Inspired Network With Automatic Brightness Adaptation for Pavement Crack Detection in Different Environments

Accurate and efficient pixel-level pavement crack detection is critically important for traffic management. Despite the improved performance on current deep learning-based detection methods, they still face challenges due to diverse detection environments, irregular crack shapes, and varied orientations. To address these challenges, we draw inspiration from the biological vision mechanism and propose a bio-inspired pavement crack detection network (PCDNet) for pixel-level detection of road cracks in different environments. Specifically, a brightness adaptation-feature processing (BA-FP) module is proposed for adapting to different environments, which simulates the brightness adaptation mechanism of the biological visual system to perform luminance adjustment and feature processing on the input image. A Gabor difference block (GD_Block) is designed for extracting crack features, which simulates the biological visual system to different directions and frequencies of visual stimuli. It enables the extraction of crack shapes and orientations while suppressing the background texture interference. Moreover, a probability-weighted fusion (PWF) module is presented for fusing different lateral outputs. It extracts effective features by probabilistically weighting and fusing lateral outputs at adjacent scales based on the probability of each crack pixel. We conducted comprehensive experimental comparisons using five publicly available datasets: BJN260, Rain365, Sun520, DeepCrack, and BRS. The results show that PCDNet is highly competitive among all methods and achieves state-of-the-art performance with fewer computational requirements. The code can be found at https://github.com/zwuser1227/PCDNet.