Research And Application Of Deep Convolutional Adversarial Network Model Based On Semi-supervised Coding

Generative Adversarial Network(GAN),as a deep learning model,is gradually being used in supervised and semi-supervised image recognition by virtue of its advantages in extracting real images and generating image rich features in the training phase.Focusing on the improvement of the image recognition accuracy of the GAN model,the following work has been done on the basis of the original GAN model:(1)The original GAN model depends on the learning of the discriminator to a certain extent on the image features,which leads to a certain offset between the extracted data features and the real data features,which affects the accuracy of GAN model image recognition.In response to this problem,this article adds an encoder structure to the original GAN model,extracts feature from real data,and increases the weight of real data to improve the offset problem between data.(2)The training of the original GAN model requires a large amount of label data.In real life,the label data is often difficult to obtain or the number is small,and it cannot provide enough training data for the model.In response to this problem,this paper introduces a semi-supervised method into the improved model described in(1),Semisupervised adversarial training is carried out to improve the accuracy of model recognition when only a small amount of label data is used.This article gives the structure of the imported semi-supervised mechanism model,and describes in detail the structure and interrelationship of the generator,encoder and discriminator,and explains the training process of the model.The model is verified on three data sets of MNIST,CIFAR-10 and Fashion-MNIST.The experimental results show that compared with other derivative models of the GAN model,the improved model in this paper effectively improves the accuracy of image recognition.(3)In order to further verify the effectiveness of the improved model in practical applications,medical imaging—cerebral infarction MRI images are selected for image recognition in practical applications.In the experiment,the labeled samples were set to25,50,100,250,600,1200,and a large amount of unlabeled data were trained to realize the identification of the lesion in the MRI image data of cerebral infarction.The experimental results were compared with other supervised,semi-supervised models and GAN derived model,and the results proved that the model proposed in this paper can obtain better recognition accuracy in specific applications.