Research And Application Of Image Recognition Method Based On Deep Generative Adversarial Networks

With the rapid development of computer technology,a large amount of image data has been generated.The reasonable classification of these data is not only conducive to the manage of data and provide convenience for data query,but also can improve the efficiency of mining effective information and prepare for the research in the field of computer vision.Therefore,it is very important to study an efficient and accurate image recognition method.At present,image recognition technology mainly includes traditional image recognition technology and deep learning image recognition technology.Compared with the traditional image recognition technology,the deep convolution networks can extract local and global image features adaptively according to the classification task,which has a good recognition performance.However,the image recognition method based on the deep convolution networks needs a lot of data for training.When the training samples are not enough,the effect of image recognition will be reduced.Difficult samples can guide model training,which are easy to make mistake for classification,but contain boundary information samples.Therefore,how to solve the lack of training samples and difficult samples is a problem worthy for further study.In order to solve the above problems,on the basis of conditional generative adversarial networks and deep convolutional generative adversarial networks,combined with focal loss function,an online image recognition model based on focal loss conditional deep convolutional generative adversarial networks is proposed,named F-CDCGAN.The details are as follows:(1)The self-contained classifier of the generative adversarial networks(GAN)can only deal with two-class task,which cannot be used for the task of multi-class image recognition.The F-CDCGAN model proposed in this paper,on the basis of conditional generative adversarial networks and deep convolutional generative adversarial networks,constructs a generation network which can generate sample data and a discrimination network which can distinguish the true and the false.At the same time,a classification network which can distinguish categories is added,which can be used for multi-classification task;(2)The F-CDCGAN model feeds the image data generated by the generation networkand the real samples into the discrimination network and classification network,which can solve the problem of insufficient training samples;(3)Inspired by the idea of Boosting,the F-CDCGAN model proposed in this paper introduces Focal Loss function,which increases the weight of difficult samples that are easy to make mistakes in the training process,enhances the guidance of difficult samples on optimization,and thus improves the efficiency of model training.In this paper,F-CDCGAN model was used to conduct experiments on MNIST data sets and Fashion-MNIST data sets.In order to prove the advantages of the F-CDCGAN model,a CNN model with the same structure as the F-CDCGAN model classifier was built for comparison.We also compared the proposed method with other traditional image recognition methods.The experimental results on the two data sets show that the F-CDCGAN model presented in this paper performed better.