Few-shot Learning Algorithm Research Based On Multi-scale Feature Measurement Fusion

In real life,humans can quickly build up their cognitive ability of a new concept with one or a few examples,but neural network models in deep learning usually require sufficient labelled data samples for training in order to achieve the effect of predicting unknown image classes.General image classification methods have limitations in many specific application scenarios,as some labelled data samples are very sparse or require significant costs to complete the image labelling process.In recent years,few-shot learning has become a popular research area in deep learning,which can identify and classify unknown images with only a small number of labelled data samples per category;however,each category has only a small number of data samples,and these labelled data samples are insufficient to represent the feature space of the current category.This paper proposes to use local feature information in the image,which can fully exploit the local feature information that is distinguishable in the image;at the same time,this paper also proposes to use the feature maps of different scales of the image,calculate the cosine similarity of each local feature vector in the query set image and each local feature vector in the support set image,and finally fuse the results of different feature maps.The specific research of this paper is as follows:(1)To address the problem of sparse labelled images in few-shot learning,this paper proposes the use of feature maps of different scales of images,which will be processed to obtain feature maps of different scales after the feature extraction network;the use of feature maps of different scales can make use of more effective information of the images themselves and lay the foundation for later being able to obtain classification results of feature maps of different scales.(2)This paper proposes to use all the local feature vectors in an image,calculate the cosine similarity of each local feature vector in the query image and each local feature vector in each category of images in the support set,and sum the similarity results cumulatively to obtain the similarity score results of the query image.(3)For a query image that needs to predict a category,it can get different metrics from different scale feature maps,so this paper proposes to fuse the metric results of different scale feature maps of the query image.Specifically,the different metric results are weighted and assigned to obtain the final inferred result of the image.The classification effectiveness of this algorithm was verified on a few-shot benchmark dataset and a few-shot fine-grained dataset,and experiments were conducted for different datasets for the 5-way 1-shot and 5-way 5-shot few-shot image classification tasks respectively;the algorithm proposed in this paper was able to effectively improve the accuracy of image classification compared with the classical methods in few-shot learning.