Research And Application Of Image Data Augmentation Technology Based On Generative Adversarial Networks

With the development of society,science and technology,the research and application of artificial intelligence technology has attracted much attention.In 2017,the “Development Plan for the New Generation of Artificial Intelligence” issued by The State Council has established it as a national development strategy.Machine vision is an important development branch in the field of artificial intelligence,which is one of the core technologies of artificial intelligence and plays a crucial role in its development.Machine vision technology based on deep convolutional neural network is the current research hotspot,which has very important theoretical research significance and practical application value.At present,there is a wide range of data imbalance,especially to much-needed,large-scale,high-quality labeled training data resources are scarce or not easy to obtain,which has become the bottleneck problem to be solved.Using Generative Adversarial Networks(GANs)to extract the features of original data set and generate new target domain images has become the preferred method of many scholars in the research of data augmentation technology.How to correctly introduce high frequency detail information to improve image data quality is the key problem urgently.In view of the this,based on GAN,this paper takes the application of nanoparticles’ scanning electron microscope image super-resolution and X-ray image computer aided diagnosis of pneumoconiosis as the entry point,and faces the common problems of model instability of convergence and low synthetic quality in GAN-based data augmentation technology.Based on creatively proposed dual-staged loss driven super-resolution technology and local-global fused discriminant auxiliary disentangled image translation technology,we bring up novel image super-resolution and image translation models to deal with the essential problem of low signal-to-noise ratio of generated images,incorrectly embedding of high frequency details,and the insufficient style transfer ability of target region.The main innovation works are as follows:1.A Dual-Staged Loss Driven Network(DSLDN)is proposed to improve the image super-resolution quality.The model is divided into two stages: In the first stage,the deep residual channel attention network and the Mean Absolute Error(MAE)loss function are used to improve the Peak Signal-to-Noise Ratio(PSNR)of the generated images.In Set5 test set,PSNR increased from 28.417 to 32.139;In the second stage,discriminator network and adversarial loss function are used to maintain a high level of PSNR,and high-frquency details are introduced to improve the quality of image data.In SET test,Structure Similarity Index Measure(SSIM)and Information Fidelity Criterion(IFC)can reach the highest level,respectively:0.894 and 3.709,which are improved by 9.4% and 34.7% compared with non-deep learning methods.Both visual and quantitative evaluation are superior to the current same kind of models.2.A Local Discriminant Auxiliary Disentangled Network(LDADN)is proposed to improve the quality of image translation.Based on the classic style transfer methods,we originally propose local-global fused discriminant auxiliary disentangled image translation technology,with the introduction of multiple independent local discriminators and local adversarial loss function,to generate different high frequency details in the corresponding local patch of the target domain images.It differentially realizes the correct embedding of high frequency information,that is global and multi-local loss fused to generate high quality images.The test results show that the performance of Inception Score(IS),Fréchet Inception Distance(FID),Kernel Inception Distance(KID)and Learned Perceptual Image Patch Similarity(LPIPS)has a great improvement compared with the current mainstream models,reaching 1.460,1.327,5.513±1.152 and 0.512±0.076,respectively.3.A new method of nanoparticles scanning electron microscope image superresolution based on DSLDN model is presented.An experimental training dataset(120,000)was established;The loss function of the model is optimized;The model parameters are optimized;The concept of “Catalytic Aactive Site” is put forward,and a method of super-resolution quality evaluation of electron microscopy images based on catalytic active site is established.The effectiveness of the methods is verified by chemical catalytic experiments.The results of the widely used image super-resolution quality quantitative evaluation method in the literature and the catalytic active site evaluation method proposed in this paper both show that more detailed information of nanomaterials can be obtained from the super-resolved images by using this method.It provides a new tool for the efficiency analysis of cataysts with active sites.4.A new augmented X-ray image computer aided diagnosis method for pneumoconiosis based on LDADN model is presented.The original data sets(health: 1528,pneumoconiosis: 904)were established by us and our cooperations.The network model is optimized,and the local Laplacian filters and loss function are introduced to transfer the details of the focal area of the image.After training the classifiers,the classification accuracy is improved from 92.40% in the original dataset to 99.31% in the augmented dataset,which exceeded the current accuracy of human experts and effectively improved the performance of the computer-aided diagnosis system.