Multi-view Road Disease Detection Based on Attention Fusion and Distillation

The radar dataset collected by the threedimensional ground- penetrating radar is presented as multiple views, which is difficult to analyze manually. Disease detection based on multi-view radar maps extremely requires expert experience and knowledge. The high cost of labeling results in a small number of samples, which makes the task more difficult. One solution to this problem is to create a deeper network to extract disease features, but this is not conducive to practical use. Therefore, we propose a two-stage attention fusion and distillation model for multi-view road disease detection, which enables us to make full use of multi-view datasets and improve their practical application in road detection. Experiments show that our model can use fewer parameters and calculations to achieve high accuracy on both original and enhanced datasets.