Research On Precise Image Retrieval Method Based On Generative Adversarial Networks

With the development of the Internet,transmission capabilities and storage capabilities of network device,more and more users choose to share and download various image data on the network.These image data have exploded exponentially in recent years,but such large-scale image data without accurate annotation has brought a big challenge to the corresponding image retrieval tasks.Therefore,how to perform fast and accurate image retrieval on the images dataset has become a hot issue in current research.The DCNN(Deep Convolutional Neural Network)depth model is also widely used to deal with image retrieval problems because of its excellent performance on computer vision problems.The "deep" models used in the existing image retrieval methods mostly obtained good result by using supervised learning to train the models,while the models need to have a large number of labeled image data sets to train.This obviously limits the scope of application of the method,because in the actual image retrieval application,obtaining a large amount of annotation data required for training the model is very difficult and costly.For solving the above problems,this paper proposes an image retrieval model based on unsupervised training,which is based on Generative Adversarial Networks(GAN)as the basic model and improves it.It learns through the unsupervised adversarial training of the model.The image index is then applied to the image set for accurate retrieval,so that the model can be applied in the unlabeled data set.The main work of this paper is as follows:1.Propose and use the InfoGAN as a corrector of the hash network to generate more accurate image index for image retrieval.2.The performance of the original model highly depends on the length of the information vector.If the length is set inappropriate,the model training will collapse or the training result will be unstable,which cause the bad result of the model in the end.In response to this problem,the adjustment vector and adaptive information loss are introduced innovatively in this paper to eliminate the influence of information vector length on model training.3.The problem of the instability of the adversarial training process and the mode collapse caused by the unstable adversarial training process.After introducing the BN module(Batch Normalize)and the self-attention(self-attention)in this paper,above problems can be stopped and the performance of the model is promoted.Finally,in order to test the performance of the unsupervised image retrieval model based on the GAN proposed in this paper,in the experiment,the average performance accuracy(mAP)of the retrieval performance is used as the performance evaluation,and test on the dataset such as cifar-10 that remove labels already.Compared with other traditional and unsupervised methods and some methods based on "deep" model,the experimental results show that our model is superior to other models in unsupervised retrieval accuracy.In addition,in order to better verify the effectiveness of the proposed model on complex feature extraction,the model is applied to the task of face similarity measurement.And the experimental results show that the similarity detection rate is improved by rearrangement of the model.