The latest advances in deep neural networks demonstrate their ability to learn visual models of large-scale datasets. However

there is currently no publicly available dataset for road damage detection in China. The data used in this experiment is captured by a bicycle recorder. The image resolution is severely affected by vehicle occlusion

lighting changes

vehicle shaking

and other factors

resulting in a huge amount of data. However

manually selecting qualified data is time-consuming and laborious. This article proposes a new detection method that uses the feature transfer network (FTN) to reconstruct road damage images

sharpens image features while using a small amount of data

and then uses the Faster Regions with CNN features (Faster RCNN) network for damage detection. The experimental results show that using feature style transfer to reconstruct images for damage detection improves the average accuracy of different road damage category instances by 2%

demonstrating the effectiveness of the framework.