N-pair Center Loss With Feature Fusion For Fine-Grained Image Retrieval

Image retrieval tasks tend to be refined with the development of image retrieval,and research on fine-grained image retrieval has become one of the current research hotspots in the field of visual retrieval.Most existing fine-grained image retrieval schemes are built based upon deep feature learning paradigms,which typically leverage the feature maps of the last convolutional layer as features.However,such representation focuses only on the global information of the object,leaving the local details unexploited,which is however crucial to identifying subtle differences for fine-grained retrieval.To this end,this paper designs a fine-grained retrieval method that combines multi-scale deep features.By fusing the deep features of different scales,we get more discriminative features.In order to reduce the complexity of constructing tuples and improve training efficiency,this paper proposes a metric learning method based on N-pair center,which improves the efficiency of tuple construction and the ability to distinguish features by effectively optimizing the inter-class distance and inter-class distance of samples.The main work of this paper is as follows:1.This paper proposes a fine-grained image retrieval method based on multi-scale feature fusion.This method obtains the features of various scales by output feature maps at different levels of the neural network,then aggregates the features of different scales through generalized means pooling,and finally embeds them into the metric space for learning.The experimental results show that:the fusion feature generated by concatenating embedded features has higher retrieval accuracy.2.This paper proposes a metric learning method based on N-pair centers and a strategy for building centers based on sample means or category features,thereby solving the problem of too high complexity caused by difficult sample mining and improving tuple construction efficiency.By introducing the idea of "N-pair centers",the features within a class are more compact,and the features between classes are more dispersed,which is helpful to improve the ability to distinguish features.The research of this paper has conducted a large number of experimental verifications on multiple classic fine-grained image retrieval datasets and compared with current classic methods.Experimental results show that the fine-grained image retrieval method based on N-pair center loss and multi-scale feature fusion proposed in this paper has good retrieval accuracy and generalization.