Research And Application Of Image Recognition Method Based On Semi-supervised Deep Generative Network

This paper designs and implements a semi-supervised image recognition method based on deep convolutional neural networks and generative adversarial networks in order to solve the problems of the actual training of traditional deep neural networks such as the difficulty in obtaining labeled data,instability of model effect,and sensitivity of the model to the influence of data distribution.In the process of studying deep convolutional semi-supervised generative adversarial networks(DCSGAN)using in image recognition,this paper mainly does the following three tasks:(1)Design and implementation of image recognition model based on semi-supervised deep convolutional generative adversarial networksBy studying and learning traditional deep generative models,according to the abovementioned problems,this paper proposes a semi-supervised image recognition method using deep convolutional generative adversarial network(DCGAN).It uses the characteristics of adversarial training for generative adversarial networks to train the generator and discriminator at the same time.The training generator fits the complete data distribution according to the labeled data to achieve semi-supervised training.The data is enhanced by the generator to improve the model effect.Finally,the discriminator that has been trained is applied to the image recognition tasks(2)Design and implementation of generators and discriminators of generative adversarial networks based on Res NetTraditional deep convolutional generative adversarial networks use ordinary convolutional neural networks as generators and discriminators,which have degradation problems such as gradient explosion and gradient disappearance in the actual training process.This problem is particularly obvious in the process of increasing the depth of the network,and such a problem has a greater impact in the process of confrontation training for generating a confrontation network.Therefore,this paper uses Res Net to replace the traditional convolutional neural networks as the discriminator and uses batch normalization method the generator of the generative adversarial networks,which can reduce the degradation problem in the training process effectively,and make the model effect better by deepening the network layers.(4)Design of the discriminator adapted to image recognition tasks.The discriminator of the traditional generative adversarial networks is a two-classifier,which is mainly used to determine whether image is in the real data distribution,and it cannot determine the specific the image categories.This article uses multiple classifiers to replace traditional discriminators to meet the needs of image recognition tasks.