Hierarchical Convolutional Neural Network With Feature Preservation and Autotuned Thresholding for Crack Detection

Drone imagery is increasingly used in automated inspection for infrastructure surface defects, especially hazardous or unreachable environments. In machine vision, the key to crack detection rests with robust and accurate algorithms image processing. To this end, paper proposes a deep learning approach using hierarchical convolutional neural networks feature preservation (HCNNFP) an intercontrast iterative thresholding algorithm binarization. First, set of branch proposed, wherein output previous blocks half-sizedly concatenated current ones reduce obscuration down-sampling stage taking into account overall information loss. Next, extract map generated from enhanced HCNN, binary contrast-based autotuned (CBAT) developed at post-processing step, where patterns interest are clustered within probability identified features. The proposed technique then applied identify cracks on roads, bridges pavements. An extensive comparison existing techniques conducted various datasets subject number evaluation criteria including average F-measure ( $AF_\beta $ ) introduced here dynamic quantification performance. Experiments images, those captured by unmanned aerial vehicles inspecting monorail bridge. outperforms methods tested GAPs dataset increase about 1.4% terms while mean percentage error drops 2.2%. Such performance demonstrates merits HCNNFP architecture defect inspection.