Surface Defect Data set Enhancement method for wind Turbine based on RES-DCGAN

Abstract

In order to solve the problems of low image resolution, high sample similarity, low stability and parameter oscillation of the DCGAN model during the generation of training model. A network structure based on residual network to enhance generator and discriminaton model. Secondly, the loss function was replaced, W(Wasserstein) distance was used and spectrum normalization (SN) was introduced to improve the traditional DCGAN model, and the images generated by the improved model and the unimproved model were detected by MaskRCNN target detection algorithm. The experimental results show that the improved DCGAN model can better generate target images, give more prominence to details such as target shapes in fan surface defect areas, and effectively improve the accuracy of target detection by 7.6%.