Research On Improvement Of Classification Algorithm Based On Triple-GAN

Image classification,as one of the important means to understand the content of images,has been successfully applied in the fields of finance,public safety,and transportation.Its importance is self-evident.Faced with a large number of image data sets,the use of artificial methods to rmark the semantic attributes of images is not easy to search and categorize,resulting in problems such as inaccurate analysis of image information.With the introduction of relevant GAN frameworks,GANs has become an important model in the field of deep learning and an important tool for artificial intelligence research.And proposed a variety of improved models based on GAN framework.In particular,Triple-GAN developed the GAN framework from a two-player game to a three-person game.That is,a classifier was added on the basis of discriminators and generators,so that Triple-GAN could solve the problem that the generator and discriminator could not reach the maximum at the same time.The generator and generator cannot control the semantics of the generated sample.However,Triple-GAN still has the following problems in image classification:First,because Triple-GAN needs to mark some sample tags in the classifier,Triple-GAN still uses manual methods to mark samples,causing manual marking workload.Too large and uneven markings.Second,Triple-GAN still uses the KL divergence distribution to construct the target loss function.However,when the KL divergence does not intersect,the gradient will disappear.So similar problems exist in the GAN,such as the gradient,in the Triple-GAN.Disappearance,training instability and other issues.The main innovations of this paper will be divided into two aspects:(1)To solve the first problem in Triple-GAN,this paper proposes to improve the Triple-GAN classifier using the random forest classification algorithm.A random forest decision tree is established through random subspace and Bagging.In the process of building a decision tree,leaf nodes are automatically tagged.At the same time,it is ensured that in the training,through the predictive analysis of the random forest algorithm,each training sample can be located on the leaf node,and finally the tag pair formed by the leaf node tag and the training sample is input into the discriminator.The process of marking has changed from manual marking to automatic marking,and the classification efficiency has been significantly improved.(2)According to the second problem of Triple-GAN existence,this paper proposes the idea of LSGAN theoretical model and improves the Triple-GAN objective loss function,The target loss function is constructed by minimizing the chi-square distribution and parameter variables.The sample distribution is controlled in a stable confidence space while the direction of the generated sample is adjusted.The generated sample is always controlled on the correct side of the decision boundary so as to achieve the effect of training stability.Through the above two aspects of optimization and improvement,to build Triple-GAN improved model-Improved Triple-GAN model,using the MINIST,cifar10 and cifar100 data sets were conducted on the Improved Triple-GAN model and Triple-GAN model experiments.Experimental results show that the Improved Triple-GAN model solves the problem of manual tagging compared to the Triple-GAN model and avoids the situation where the gradient disappears and the training is unstable.As a result,the Improved Triple-GAN model is more stable than the Triple-GAN model,and the training result is more ideal.