Research Of Image Classification Based On Generative Adversarial Nets And Its Application For Pulsar Candidate Identification

Pulsars are highly magnetized,rotating neutron stars with small volume and high density.The discovery of pulsars is of great significance in the fields of physics and astronomy.With the development of artificial intelligent technology,image recognition methodologies based on deep learning are increasingly adopted for pulsar candidate identification.However,pulsar candidate dataset is characterized by unbalance and lack of positive samples,which has contributed the traditional methods to fall into poor performance and model bias.In addition,a large number of labels are required for training these methods,which is also difficult for pulsar candidate dataset.To this end,from the perspective of generative adversarial nets(GANs),this dissertation studies the problems of image recognition and semi-supervised learning for unbalanced dataset and small-scale dataset.The main results and innovations of this dissertation mainly are rendered in the following four aspects:(1)A novel image recognition model CP-ACGAN is proposed.The auxiliary classifier GANs(ACGAN)is a sample controllability generation model based on GANs,and it can predict the labels for samples.However,it shows slow convergence and poor recognition performance when the model is applied to image classification.The analysis reveals that it is caused by the network structure of the output layer of discriminator and loss function of the model.Therefore,a novel image recognition model CP-ACGAN is proposed by improving the network structure,reconstructing the loss function,and adding the weight factor between sample generation and sample classification.The generated samples in the model are employed to augment the diversity of training samples,thereby improving the image recognition effect.At the same time,the analysis of the weight factor indicates that the CP-ACGAN model is a unified expression of ACGAN,DCGAN and CNN models.Finally,the classification experiments are carried out on SVHN and CIFAR10.The results manifest that the CP-ACGAN has better recognition performance than traditional recognition methods,and experimental analysis of the weight factor in the model is also conducted.(2)A novel semi-supervised learning model SSL-ATJD is proposed.A semi-supervised learning model SSL-ATJD is presented based on the sample generation ability of GANs.The model consists of a generator,a classifier and three discriminators,and four types of joint distribution between samples and labels are involved for adversarial training.Theoretical analysis demonstrates that the model has a unique optimal solution,and the samples produced by the generator can effectively supplement the diversity of labels.When it reaches equilibrium,the four types of joint distribution are equal,and the corresponding conditional distribution and marginal distribution are also equal.Therefore,the model can not only predict labels for samples,but also controllably generate samples.Finally,semi-supervised experiments are conducted on MNIST,CIFAR10 and SVHN,and the results illustrate that the SSL-ATJD model achieves state-of-the-art semi-supervised classification results and semi-supervised generation ability;Meanwhile,further experiments on MNIST also indicate that the model is extremely robust to the number of labels in semi-supervised learning.(3)A novel image classification model ICAT with adversarial training is presented.The image recognition model ICAT,which contains a generator,a classifier and two discriminators,is obtained by further improving SSL-ATJD.In the training,the generator and the classifier supervise and cooperate with each other to achieve the optimal together.Theoretical analysis manifests that the model has a unique optimal solution.When the model reaches equilibrium,the classifier happen to be the inference network of the generator,which means that,for conditional generation samples,the prediction labels of the classifier are exactly the same as the input labels of the generator.Therefore,the samples produced by the generator effectively augment the diversity of training data.The ICAT model not only performs excellent recognition ability for both small-scale dataset and unbalanced dataset,but also can generate controllable samples.Finally,small sample experiments are conducted on MNIST and SVHN respectively.The results declare that ICAT model not only has better classification performance than CNN,but also has better controllability than CGAN and ACGAN models.(4)The proposed models are applied to pulsar candidate datasets FAST and HTRU.The image recognition models CP-ACGAN and ICAT proposed in this paper are applied to the pulsar candidate datasets FAST and HTRU,respectively,to address the problems of poor identification effect and model deviation faced by traditional deep learning recognition models.The results expose that the proposed models not only improves the F-score values,but also reduces the false negative rate.By contrast,the recognition effect of the ICAT model is better,so it is more suitable for the identification of pulsar candidates.Finally,the SSL-ATJD model is applied to HTRU to explore semi-supervised learning in unbalanced dataset.The results reveal that,with small number of labels,the model achieve the same recognition result as the fully supervised CNN models.Therefore,it can not only cut down the manpower and material costs caused by labeling,but also play a significant role in the dataset of scarce labels.