Research on Crack Detection Method of Wooden Ancient Building Based on YOLO v5

Abstract

We proposed a crack decoction method based on YOLO v5 for the purpose of increasing efficiency, cutting costs and short the time of the crack detection. Firstly, we created a Pascal VOC dataset based on the wooden ancient building photos we have taken by a Canon camera; Then we trained the dataset with YOLO v5 based on the Pytorch which was an open source machine learning framework supported by Facebook; Finally, we tested the method we proposed with a famous building named Bawang temple. The test results shows that after 350 times trained the model, the value of loss function was 0.042 and the AP value was 0.918, the precision of the model was 91%. The experimental results verify that it is feasible to analyze crack by YOLO v5 with faster and less waste than the traditional way.