A crack detection method based on deep transfer learning

The crack recognition method based on deep learning requires a huge amount of data and it is difficult to collect adequate data manually, which leads to insufficient training and poor recognition. To solve this problem, a crack recognition method based on deep transfer learning (DTL) was proposed. An unmanned aerial vehicle (UAV) was used to quickly collect and product crack data set, then convolution layers and the fully connected layer of visual geometry group-16 (VGG-16) were trained on ImageNet data set and crack data set, respectively. Finally, a DTL model was built by connecting two parts with freezing and fine-tuning some parameters. Experimental results showed that the method significantly reduced image requirements, and improved the precision of crack detection. 95.9% of the recognition precision was obtained on the validation set, and 0.983 AP value was obtained on the test set.