Research On Deep Metric Learning Algorithm Based On Hard Sample Generation

With the rapid development of information technology,people are using more and more online technologies frequently,and thus the data generated is becoming more and more complex.Traditional metric learning is currently unable to handle more complex data,and with the rapid development of deep learning,researchers are also focusing on deep learning,existing researchers focus on improving the performance of deep metric learning models,who use hard sample mining strategies to find more hard sample pairs to guide the efficient training of models.But hard sample mining methods tend to focus on hard sample pairs with large losses in part of the datasets,while ignoring easy sample pairs with small losses in the datasets,resulting in models that cannot accurately portray the global structure of the embedding space.In view of the above problems,hard sample generation methods have attracted a lot of attention from researchers.In this paper,a research work is conducted on the deep metric learning algorithm based on hard sample generation,and the main research contents are described as follows.(1)A deep metric learning method for generating hard positive samples based on sample rotation is proposed.The positive samples in the triad are rotated to the reverse extension of the line connecting the anchor point and the class center on the axis of the class to which they belong,and a new loss function is given to construct a deep metric learning model based on sample rotation to generate hard positive samples,the efficiency of the training model is effectively improved.The results show that the proposed deep learning algorithm based on sample rotation for hard positive sample generation outperforms other common hard sample generation methods for image retrieval.(2)A deep metric learning method for generating hard tuples based on sample rotation is proposed.The method takes full account of the sample distribution information when generating discriminative hard tuples,and rotates positive and negative samples to the line connecting the corresponding class center and the anchor point or its reverse extension,respectively,under the constraint of class center to generate hard tuples,and a deep metric learning model for generating hard tuples based on sample rotation is constructed,and the effectiveness of the proposed algorithm is verified by conducting experiments on image retrieval tasks on three datasets commonly used for deep metric learning.In this paper,we summarize the existing problems of hard sample generation,improve the method of generating hard samples by algebraic computation,and propose a deep metric learning method based on sample rotation to generate hard positive samples.The deep metric learning method based on sample rotation to generate hard tuples is proposed to improve the existing hard tuple construction method.A new research idea is provided for the deep metric learning algorithm.