Research And Implementation Of Augmentation Strategy For Gan Training

In the increasingly information-rich modern society,images play an irreplaceable role in conveying information.With the continuous development of deep learning technology,the effect of image generation become increasingly important and far-reaching in various fields.The generated high-quality images can make up for the insufficient sample size of the original image set,so that the trained model can achieve better accuracy and effect.Among them,for some special types of images,such as specific characters,rare animals and plants,the corresponding generation tasks suffer from small sample size and difficult acquisition,and the generative adversarial network,which plays an irreplaceable role in the field of image generation,is limited by the sample size.In the case of limited size,the performance of trained models will be greatly reduced.To deal with the problem,there are two research directions,based on transfer learning and data augmentation at present.Among them,the differentiable augmentation technology based on the data augmentation direction has achieved good results,but its data augmentation strategy is fixed.It does not fit well with different types of datasets.Based on the differentiable augmentation technology,this paper proposes a training method for automatically searching for data augmentation strategies without changes to loss functions or network architectures.Through experimental comparison,the performance of the trained generative adversarial network model is significantly improved.Combined with this method,a small sample image generation system based on generative adversarial network has been designed and implemented.The specific contents of this paper are as follows:Combined with differentiable augmentation technology and the data augmentation strategy automatic search PBA algorithm which has achieved remarkable results in the field of image recognition,a data augmentation strategy automatic search algorithm PBAG suitable for generative adversarial network training is proposed.Experiments show that the model trained by this algorithm achieves better results than the training method of manually selecting the augmentation strategy and the small-sample image generation method based on transfer learning,and can generate higher-quality images.Additionally,aiming at the high computer configuration requirements of the neural network training process,the cumbersome tuning process and the need for visualization of data,a small-sample image generation system based on a generative adversarial network with a B/S architecture is designed and implemented,to meet the needs of deep learning researchers to expand small-sample image datasets and manage content related to image generation tasks.