Research On Few-shot Image Classification Algorithm Based On Task-relevance

Depth learning algorithm related to image classification uses a large number of labeled images to train the depth network model,and makes the model better predict the unknown image categories by continuously optimizing the model parameters.However,for some scenes where the sample data is difficult to obtain and the image annotation needs to be completed at a high cost,the traditional image classification algorithm is not applicable.Therefore,the research on the method of training depth model in few-shot scenario will help to reduce the research cost and expand the application scenario of deep learning.In order to improve the accuracy of few-shot image classification,this paper uses the few-shot learning task training mode to learn different task structures in the training data to adjust the sample characteristics,and optimize the method of solving the image prototype,so as to improve the accuracy of image classification.The main research contents are as follows:(1)In view of the neglect of the correlation between tasks in the traditional fewshot learning image classification,a task correlation model algorithm is proposed to further mine the few-shot learning task information.By learning the relationship between the data in each training task,the model generates a mask with the same length as the sample feature vector to adjust the data features,so as to better adapt to different classification tasks.When testing new class data,the module can also be used to better adapt to classification tasks and improve generalization.The experimental verification of the model on Mini Image Net dataset proves the effectiveness of this method.(2)Aiming at the deficiency of the simple average idea proposed by the prototype network to solve the prototype,a prototype correction module algorithm is proposed.The module includes two strategic schemes.Firstly,because different samples have different contributions to the prototype,the concept of sample weight is proposed to distinguish samples by different weight values.The input of the weight network is the support set in the few-shot learning tasks,and the output is the weight value of each sample,and then the prototype is obtained by weighted summation.Secondly,the unlabeled data is introduced to enhance the support set data,and the query set samples are continuously screened by using appropriate strategies,and then merged into the support set,so as to enhance the reliability of the prototype solution on the data samples.Experiments show that the module can promote image classification.At the same time,the module is fused with task related modules.The experiments show that the effect of the combination is further strengthened.