Research On Cracked Insulator Image Generation By Generative Adversarial Nets

Deep Learning had broken down bottlenecks with development of graphic hardware,drone smart patrol technology is growing fast,and it includes the use of deep learning to implement intelligent detection of aerial defect insulator.Due to the special nature of various power equipment,it is difficult for professional technical personnel to conduct on-site surveys and inspections of power equipment with low efficiency.Object detection model based on deep learning could identify location quickly,but it needs a balanced cracked insulator sets to solve low recall.In fact,it’s difficult to get such insulator with abundant fault sample in each kind.Generative Adversarial Nets provide a possible method for the balance of the training insulator base.Building an insulator image enlargement model byGAN,it could be used for enriching defect insulator image sets for training of defect insulator detection model.Get contour image of insulator by Holistically-Nested Edge Detection algorithm,pair insulator and contour image in order.,building ContourGAN model for the generation of contour and insulator image generation.Attention mechanism make progress in visual detection and image segmentation,the use of attention mechanism make it possible to improve the quality of generated image.Build AttentionGAN model to generate cracked insulator image.From the result of generated insulator image,those two methods could not generate unseen and expected insulator image which has different fault location compared with original insulator.CycleGAN is a Generative Adversarial Networks that based on image style transfer,and it make success in the translation between orange and apple each other,but it met trouble in transfer such thin and long object like insulator in image.Use a modified CycleGAN model for the enlargement of cracked insulator sample base.Prepare two-value mask image of insulator before experiment,two-value mask image divide image into global or local insulator area,where the purpose of global insulator mask is that image translation between different insulator texture,or different cracked location feature generation for local area.Enlarge insulators datasets by improved CycleGAN and modify loss function for the purpose of better model,add hadamard product of generated mask image and RGB image as content loss,the replacement of loss function keep the contour of generated image,achieve part cracked insulator generation.Use Frechet distance(FID)to estimate performance of generated image and prove proposed CycleGAN model is better than otherGAN models.The datasets composed of generated image by modified CycleGAN is compared with other augmentated datasets in CNN image classifier,the experiment showed that the detection accuracy improve evidently in cracked insulator which trained in augmentated datasets by imporved CycleGAN.Modified CycleGAN could augment insulator datasets by generating diverse images with given input images,which thus could be emploved to improve other recognition or detection tasks.