Research And Application Of Transfer Learning Method Based On Deep Learning

With the rapid development of Internet technology,images,video and other data resources on the network have shown explosive growth.People can take photos and upload them to the Internet anytime and anywhere through WeChat,QQ,Facebook and other social tools.However,a key weakness is that most of the data is unannotated.Faced with such massive unlabeled data,how to let the computer automatically understand the content,and then carry out effective classification,management,retrieval and recommendation,etc.,has become a problem that the academia and industry urgently need to solve.Unlike humans,who are very good at identifying rarely seen objects without direct supervision or using knowledge in other fields to assist in identifying objects,most reliable machine learning recognition algorithms are based on a large amount of supervision information.However,annotating data is often expensive and time-consuming.To meet this challenge,knowledge on annotated datasets can be transferred to unannotated datasets for model training,which can effectively alleviate the data dependence.However,two differences between data sets make migration learning difficult: sample distribution differences and category space differences.Aiming at these two differences,this paper proposes the algorithms based on domain adaptation and zero-shot learning.The main work of the paper is as follows:(1)For the sample distribution differences: we proposed a novel architecture,which we call Attribute-assisted Networks(ASN),to transfer the knowledge from real images to sketch.Our model learns a private subspace for the real image assisted by the attribute constraint,which captures domain specific properties,such as “colors” and “fur” of animals.A shared subspace is also learned using domain constraint to capture representations shared by the images and sketches,such as edge and low level geometric.By separating the shared features from the private features of the real images,our model can separate the information that is unique to real images and produce representations that are easier for transferring to the sketch domain.Extensive experiments demonstrate the effectiveness of the ASN model.(2)For category space differences: we proposed a Semantic Attention-based Compare Network(SACN).Related prior zero-shot work maps the visual features and semantic attribute features into a subspace,then uses fixed pre-specified distance metrics such as Euclidean or cosine distance to perform classification.Our model uses a discriminative network to extract visual features,utilizes an attention-based CNN to compute a flexible distance metrics in a data driven way.Extensive experiments on two benchmarks demonstrate the effectiveness of the proposed SACN.