Target Recognition With Generative Adversarial Networks

Currently,target recognition is considered as a research hotspot in the field of computer vision.With the advantage of high recognition accuracy,the target recognition method based on deep learning has become an important technical means for target recognition.As a data-driven learning method,for deep learning,the problem related to small samples is more prominent.Therefore,this paper mainly studied the issue of small samples target recognition based on generative adversarial networks.First of all,a generative adversarial networks based on sub-feature space is proposed to expand dataset.And then,a training method of generative adversarial networks based on feedback regulation is proposed to improve the stability of training.Finally,it applied the above methods to the typical ship target recognition task.The main contents are as follows:1.Applied the typical methods of generative adversarial networks to augment the original samples data.Used the recognition networks,and conducted the classification training for both the augmented data and the data before augmentation.Through the controlled experiment,we analyzed the performance of generative adversarial networks.2.For the poor quality problem of large and complex images generated by generative adversarial networks,a method based on feature is proposed.Firstly,mapped the original small samples to a sub-feature space,and then,split and reorganized the feature data of the sub-feature space,finally generate the features of the sub-feature space by generative adversarial networks.The experiment verified that the application of this method contributed to improve the performance of the target recognition network.3.Focusing on the problem of stability insufficient in the generative adversarial networks training,a training method of feedback is proposed.This method can regulate the update frequency to generate and discrimination networks in each iterative process,by that to ensure the performance balance of the two networks and prevent the gradient vanishing.Finally,the experiment validated that the feedback regulation method could improve the performance of the target recognition networks.4.A target recognition prototype platform based on generative adversarial networks is designed.The prototype platform can display all the results of this paper.