Research On Few-shot Learning Method Based On Deep Semantic Dependencies

Deep learning relies on the mass of the labelled sample data has achieved great success,but in the case of less sample data,the performance of deep learning is not very desirable.In reality,the cost of labelling the samples is often huge.Therefore,how to identify new categories from a small number of training samples becomes the focus of research.At present,metric learning is widely used because of its simple and efficient characterist ics.The process of metric learning is mainly to construct projection space and calculation of similarity.This paper proposes a new feature extraction network based on metric learning,and uses the potential semantic information between samples to modify the position of features in the projection space.Finally,the performance of the classification is improved.The main research contents of this paper are as follows:1.Aiming at the problem that the current shal ow feature extraction network cannot represent the sample features wel,a bidirectional feature network is proposed.The network uses the residual structure and bidirectional calculation to build a deep network model.The construction of multi-projection space enables the calculation of similarity to integrate sample multi-scale information,reducing the difficulty of updating the network model and enhancing the performance of the network.Aiming at the problem that only the visual correlation is calculated in the current metric learning,the semantic saliency feature is missing.There are drawbacks in complex scenes such as background interference,a semantic network is proposed.The network first extracts the multi-sca le features of the samples.Use this semantic information to make the feature representatio ns of different samples far away.The model can learn the distinguishable semantic features of the samples and improve the accuracy of sample classification.2.Semantic information serves as the sample's attention,indicating the differe nt importance of the original sample in different regions.The model magnifies the weight of the sample on semantic information and ignores the feature interference of irreleva nt information.Correct the position of the samples in the projection space,the feature representation in the projection space is more distinguishable.The experimental results on Omniglot and mini Image Net datasets show that the proposed method can effective ly extract the semantic relationships between sample pairs to improve the classifica t io n accuracy in few shot learning.