Enhancement of pavement disease image data set based on improved GAN
The pavement disease data has rich spatial information and high correlation of feature information,while the existing image data enhancement methods are difficult to make the model capture such information ef-fectively.In this paper,an improved generative countermeasure network(CLSGAN)is proposed for pavement disease data enhancement.First,in order to avoid pattern collapse and ensure the diversity of generated images,a new loss term is introduced to reconstruct the objective function of the generator.Secondly,by integrating the coding structure of CAE,the model can obtain the spatial potential information of real pictures,which is used to strengthen the generator's learning of image spatial information,and improve the convergence speed and genera-tion quality of the model.Finally,a lightweight residual projection expansion projection expansion module(RPEPX)is constructed and spectral normalization is introduced to further improve the quality of the generated image and ensure the stability of the model during training.The experiment was carried out on the newly estab-lished CSGP data set,and the results showed that CLSGAN had a greater improvement in the evaluation indexes FID,SSIM,PSNR of the generated pavement crack and depression images.Finally,Yolov5s detection network is used to verify the progressiveness of this method.The results show that in the case of small sample data sets,compared with traditional data enhancement methods,this method achieves the optimal detection results.
data enhancementGANencoderspectral normalizationRPEPX
NETL
NSTL
万方数据
