Research On Image Data Augmentation Method Based On Generative Adversary Network

Many methods have been derived from the development of traditional data enhancement,but since they are always transformed on the original image,their improvement in classification performance is still very limited.Even if AutoAugment and other automated search data enhancement strategies appeared later,it only simplified the strategy selection The process does not substantially improve the classification performance of the data set.With the rise of generative adversarial networks(GAN),GANs can be combined according to the high-dimensional features of existing data sets to generate completely different images from the original data set.Compared with traditional data enhancement methods,it can provide more There are many image features,and a carefully designed generative adversarial network can generate images that are similar and clear to the original image,but at the same time,GAN also has unstable training,low image quality,unable to generate specified label images,and generated noise data Problems,so this paper proposes a data augmentation method based on generative adversarial networks for these problems.The main research results are as follows:1.Aiming at the problems of unstable general GAN training,low quality of generated images,and inability to specify the category of generated images,combined with the advantages of WGAN-GP and ACGAN,a generative adversarial network model DAGAN-GP was designed,combining gradient penalty A new loss function is designed.At the same time,a series of optimization strategies were introduced:label smoothing,minibatch discrimination,spectral normalization,and residual block.By comparing with WGAN-GP and ACGAN,DAGAN-GP can not only produce clearer and more distinguishable images visually,but also achieve better results on GAN evaluation indicators.On the CIFAR10 data set,compared with WGAN-GP and ACGAN,the IS index of DAGAN-GP has increased by about 12.3%and 15.1%,and the FID index has been reduced by 12.4%and 16.6%,respectively.On the S VHN data set,compared to WGAN-GP and ACGAN,the IS index of DAGAN-GP has increased by about 5.5%and 6.8%,and the FID index has decreased by 23.3%and 42.9%,respectively.At the same time,a control group was set up and a classification network TestNet was designed to compare the data enhancement effects of each method.The results prove that the method of using DAGAN-GP to generate images for data enhancement proposed in this paper has a greater improvement in classification performance.2.Aiming at the problem that DAGAN-GP still produces some noise images and affects the classification performance,the ENN-IHT data selection algorithm is proposed.First,the original data set is put into the modified VGG19 model for training to obtain the feature extraction model,and then Extract the features of the DAGAN-GP generated data set,and finally use the ENN-IHT data selection algorithm to perform data selection on the extracted feature data,remove the noise data as much as possible,so that the data retained in the generated data set can improve the classification performance Contributing data.After setting the experimental control group for verification,the results prove that the ENN-IHT data selection algorithm can further improve the quality of the data set and the performance of model classification.The image data enhancement method based on the generation of confrontation networks and data selection proposed in this paper can generate high-quality image data for different data sets,which can be effectively applied to the field of image recognition,effectively expand the image data sets,and improve the model classification performance.In the future,we will continue to optimize the network structure to enable it to run more efficiently,further adapt to unbalanced data sets,and optimize data selection algorithms to enable adaptive parameter selection.