Research On Image Recognition Method Based On Generative Adversarial Networks

With the development of modem social economy and technology,more and more information is generated in daily life and work,and the image accounts for a large proportion.How to effectively process,analyze and understand images is crucial.As one of the research hotspots in the field of computer vision,image recognition has very important theoretical significance and practical application value.Traditional image recognition methods require experienced researchers to manually select image features and then classify them by statistical classifiers,often with low recognition accuracy.In the era of rapid development of various technologies,the recognition accuracy and recognition rate of traditional methods have been unable to meet the actual application requirements.The various deep learning models represented by convolutional neural networks simulate the characteristics of human brain auto-leaming image samples through multi-level convolutional and pooling layers,and obtain classifiers whose recognition accuracy is much higher than traditional methods.GAN is a popular generation model in recent years.This paper studies image recognition methods based on generating confrontation networks.For the problem of effectively improving the recognition accuracy,this paper starts with improving the training sample set and studies the image recognition method based on the generative adversarial networks.In this paper,the Image Recognition-Deep Convolutional Generative Adversarial Networks model is proposed,and the number and diversity of training sample sets are expanded by using the generated samples,so that the discriminator network leams more image sample features in training,thereby improving the accuracy of classifier recognition.Perform image recognition experiments on MNIST datasets and CIFAR-10 datasets and CLP datasets.The results are compared with other methods.The experimental results show that the proposed method can effectively improve the recognition accuracy.In order to solve the problem that the accuracy rate of image recognition method with supervised learning is low when there are few label samples.this paper combines semi-supervised learning theory with Deep Convolutional Generative Adversarial Networks model to establish a Semi-Supervised Deep Convolutional Generative Adversarial Networks.The model uses a small number of labeled samples and more unlabeled samples for semi-supervised learning,and replaces the discriminator with a multi-classifier.Learning the overall distribution of labeled and unlabeled samples generates image samples and inputs and trains the classifier so that the classifier can have enough training samples to get an accurate classification decision surface.After the training is completed,the classifier is extracted and adjusted to obtain a network structure for image recognition.Image recognition experiments are carried out on MINIST dataset,CIAR-10 dataset and MURA dataset.The results are compared with other methods which use a large number of labeled samples.The experimental results show that this method can achieve accurate images recognition with only a small number of labeled samples.