Research On Generative Adversarial Networks For Quality Inspection

Most of the industrial inspections are based on image processing algorithm so far.There is a tendency for industries to apply deep learning to visual inspection.However,it's hard to build database from industrial domain because of samples of some defects are extremely rare and the defects on images are not obvious,which makes deep learning model difficult to converge while training.Therefore,this paper proposed an industrial product's quality inspection method based on the Generative Adversarial Networks(GAN)for commutator sample augmentation.Investigates are as follows:1.Provided a product quality inspection method framework with the generative model and the classification model using CNN for defect prediction and GAN for product data augmentation.Dissect structure of AlexNet,DAGAN and DCGAN.2.Analyzed problems encountered in the quality inspection of commutator,proposed the data augmentation model named Neighborhood Differential Filter GAN(NDF-GAN),and described its theory,structure and loss function.3.Theoretical analysis of loss functions designed in Vanilla GAN,WGAN,LSGAN and DRAGAN.Reveal how these loss functions affect model training with their optimizing goal and discriminate probability.4.Acquired and preprocessed real commutator samples before experiment.Then augmented commutator samples by NDF-GAN and training AlexNet with them under different parameters and loss functions.Compare NDF-GAN's performance with other GANs.The proposed model can converge with less real data and generate high quality samples over experiment,which exceed the performance of existent generative model.The generated samples can improve the defect prediction model's performance.Parameters and loss functions affect model's performance.In summary,this paper proposed a new method for product's quality inspection and verified it.