Research Of Image Data Enhancement Method Based On Generative Adversarial Network

With the advent of the era of big data,the demand for data has led to explosive growth,and image data is an important way of expression.Although the total amount of image data generated on the Internet is growing and the way to obtain image data is convenient,the quality of image data is not always guaranteed.For example,the number of image data samples is small,the sample categories are rare,and the sample categories are unbalanced,which will bring difficulties to the data generation.At present,the generation of image data still relies heavily on human factors,resulting in it's expensive and inefficient.Data enhancement technology for high-dimensional image data,although the traditional method can increase the number of image samples to a certain extent,large-scale generation of image samples will increase the risk of over fitting,often with limited effect.Inspired by the recent achievements of the generative adversarial network(GAN),this paper proposes two data enhancement methods.The main research work is as follows:Firstly,the traditional methods of data enhancement are described in detail,the basic theories of convolutional neural network(CNN)and generative adversarial network and the research difficulties in the field of image data enhancement are studied.Secondly,a method of data enhancement based on supervised learning is proposed for a conditional self-attention generation adversarial network.Aiming at the problem of generating image of specified category,the network model is inspired by the supervision idea of conditions generative adversarial network(CGAN).Condition variables are introduced into the generator and discriminator,and additional information is used to guide the data generation process.In addition,the self-attention mechanism is introduced based on the advantages of self-attention generative adversarial network(SAGAN).In view of the performance and convergence speed of the network,the L1 loss function is introduced to measure the pixel level difference loss of the image,which makes the network pay attention to the image feature information as well as the reconstruction of image pixel information.Then,a data enhancement method based on transfer learning is proposed for the spatial attention star generative adversarial network.To solve the problem of image detail information transformation,the network adds spatial attention mechanism to the generator,which makes the input image and the target domain have correlation,and keeps the domain independent region.In order to solve the problem of training instability and gradient disappearance,Wasserstein function is introduced into the loss function as a regular term to accelerate the convergence of network training.Finally,experiments are performed on four data sets.The image quality evaluation criteria are used to quantitatively analyze the generated samples,and experiments are performed to compare the recognition rate of the generated samples on the classifier,whichverifies the effectiveness and feasibility of the proposed method.