Research Of New GAN Models For Small Datasets And Its Application Research

In recent years,Generative Adversarial Networks(GANs)have made significant progress.However,in some fields,acquiring a dataset can be extremely difficult and costly.Additionally,GAN performance is also limited when the dataset is very small.Therefore,solving the limitation of dataset size in GAN training is of utmost research value.In order to improve the performance of GAN in small-scale training dataset scenarios,many methods have been adopted,including the concept of transfer learning and data augmentation.Although these methods have shown some effectiveness,they still face many challenges.One is that during the process of data enhancement,there is often a problem of sample noise infiltration,resulting in the generated image being mixed with additional noise information.The problem of sample noise penetration specifically refers to the fact that after introducing data augmentation,the original data distribution no longer follows the distribution of the real dataset,but rather follows the distribution of the enhanced dataset.This leads to the image data generated by the generator being biased towards the enhanced dataset,resulting in noise that does not match the real data.The false images generated by the generator should only contain sample information from real data as much as possible,and allow for noise information.However,excessive or obvious noise information may prevent the generated false images from being applied to downstream tasks.In order to train more robust models,people often add some enhanced information to the dataset samples,such as flipping,scaling,twisting,and so on.High quality enhanced information should have high accuracy,ensuring that after adding enhanced information,it can still meet the authenticity requirements of specific projects for dataset samples.When the samples generated by the generator contain noise,it is possible that the noise is caused by poor generator generation performance,or it may be due to the introduction of additional noise after data augmentation on the dataset.The training images generated under different training conditions can be compared to determine the source of noise,such as using different hyperparameter,network structures,loss function,etc.when training GAN.If the images generated under different training conditions have similar noise information,it can be determined that the noise introduced is due to data augmentation operations on the dataset samples.The problem of sample noise penetration is an additional problem caused by the introduction of data enhancement operations to alleviate the overfitting problem,so the problem of sample noise penetration can still be understood as dealing with the overfitting problem.The second is that the discriminator is prone to overfitting in the scene of small-scale training data set,resulting in the deviation of the generated image data distribution from the real data distribution.This article delves into new methods for GAN modeling in small dataset scenarios in response to the aforementioned challenges,and conducts relevant application research on this basis.The main work of this article can be summarized in the following three aspects:(1)A network model called DDGAN(Double Discriminator GAN)is proposed based on the Ada FM(Adaptive Filter Modulation)model to address the problem of sample noise infiltration caused by the mixing of generated images with the enhanced dataset during the data enhancement process.This model combines the ideas of data augmentation and dual discriminator fusion.Unlike traditional GAN models,this model introduces two discriminators to reduce the interference of non real datasets introduced during data augmentation on the generator.At the same time,the average weighting method is used to integrate the return values of the two discriminators before backpropagation,in order to adjust the weights of different discriminators and reduce the impact of non real datasets on the generated false images.Through experimental verification,it has been shown that DDGAN can effectively improve the performance of generating adversarial networks compared to the original Ada FM model,and can generate high-quality false images on 8000 datasets and 25 datasets.(2)In order to alleviate the imbalance between generator and discriminator in the training process of GAN in small dataset scenarios,the CSDGAN(Constrict Discriminator Generative Adversarial Nets)model is proposed.Based on Ada FM,this model introduces discriminator capability suppression module to reduce the probability of discriminator overfitting,thus indirectly reducing the difference between discriminator’s ability to identify true and false images and generator’s ability to generate false images in the training process.Even without the use of data augmentation,CSDGAN can still generate high-quality false images during training.Experimental research has shown that CSDGAN can generate higher quality images on smaller datasets compared to several classic models.(3)Based on the above two works,further apply the research results to practical downstream tasks to verify the effectiveness of the proposed method.In order to take into account the advantages of both methods simultaneously,this article fuses the two methods into a new network.Then,the fused network is used to generate false images and applied to two downstream tasks: image classification experiment and semantic segmentation experiment.By comparing the results obtained from images generated based on GAN with those obtained from real datasets,the effectiveness of the method proposed in this paper was effectively verified.