Research Of Transfer Learning Method Based On Sketch Labeling And Generative Adversarial Model

In recent years,deep learning algorithms have been widely used in various fields.However,the large-scale data with label is a prerequisite for good performance of such algorithms.It is well known that the cost of labeling large amounts of data is extremely expensive,and thus this precondition is not easily met.Fortunately,there are a large number of datasets that have been labeled and related to the data in the target task.Then,how to use the existing labeled data to establish a deep learning model to reduce the cost of labeling data in the case where the data has no label information in the target task becomes a hot topic of current research.Aiming at the problem of insufficient data labels in target domain,considering that generating antagonistic network can generate real distributed data,this network is extended to the field of migration learning.The labeled source domain data is used as the input of generator network and the unlabeled target domain data as the input of discriminator network.The target task is accomplished by generating the output of resisting loss and edge structure loss control generator network designed.Specific contributions are as follows:(1)This paper proposes a transfer learning method based on sketch annotation information and generative adversarial networks.Firstly,the edge annotation map obtained from the Primal Sketch map is used to train the end-to-end edge segmentation network based on the sketch annotation information,which eliminates the cost of manually labeling the image edge labels.Secondly,using the image edge information extracted by the edge segmentation network,the edge structure loss model is designed.The source domain image and its output image passing through the generator network are structurally constrained,so that the generator network output and the source domain image structure are similar.The target domain data is classified by matching the samples of the target domain data distribution.The experimental results show that the method effectively controls the output of the generator network and achieves better results compared with several migration learning methods on the digital datasets.(2)This paper proposes a generative adversarial networks based on target domain sample selection and sketch annotation.The generative adversarial networks based on the sketch annotation information does not take into account the structural similarity between the source domain data and the target domain data,so we improve the network.Firstly,the target domain samples are filtered based on the matrix norm and the pseudo-label information is given.Secondly,using the extracted edge structure information,the structural constraints between the output image of the generator network and the corresponding category image of the target domain are increased,so that the generator network output structure are not only similar to the source domain image,but also more similar to the target domain image,and the distribution is similar too;thereby classifying the target domain data.The experimental results show that the proposed method achieves higher accuracy than the transfer learning method based on the sketch annotation information and generative adversarial networks on the digital dataset,and has achieved better results than other methods.(3)The optimization training is based on the target domain sample selection and the sketch annotation generation network,and is applied to the constructed aircraft template dataset.Considering the similarity between the shape of the source domain and the target domain in the aircraft template dataset,the problem of less data samples in the target domain is achieved by adjusting the training mode of the network and adopting the pre-training method to achieve cross-domain recommendation of aircraft targets.