Research On Few-Shot Classification Algorithm Based On Enhance Feature Representation

In recent years,the rapid development of artificial intelligence can be perceived,and the convenience brought by its powerful function can be felt in our daily life and many fields.Among them,image classification is a classic research direction in artificial intelligence,with the help of deep learning algorithm,it shows very excellent performance under the support of large-scale labeled data sets.However,the current deep learning algorithm has some drawbacks,such as the need to use a large number of marked samples for iterative training.In many cases,it is difficult and costly to obtain a large number of labeled samples,such as samples of endangered animals,rare diseases,space planets,in this case,it is difficult to obtain a large number of samples for training.Therefore,the concept of few shot learning comes into being,its goal is to quickly learn the concept of new classes from a small number of labeled samples,so as to complete the identification of new classes.With the development of few shot learning,meta-learning has become a popular few shot learning framework,whose role is to develop few shot classification task models that can quickly adapt to limited data and low computational cost.In order to better solve the problems in few shot learning,this paper proposes the following two few shot learning methods based on lifting feature representation,and the specific content is as follows:First,few shot learning method based on meta-learning and location information.Recent studies on attention have demonstrated that channel attention improves feature extraction to some extent,but it ignores the role of location information,which is important for better learning from limited data in few shot tasks.Based on this fact,a new method is proposed to effectively combine location information and extracted features by pre-training a classifier with location information attention on all base classes,and then conducting meta-learning on a few shot classification algorithm based on the nearest centroid.By experiments on different data sets,competed with the current mainstream few shot image classification methods,this method has certain performance improvement on the data sets commonly used in few shot learning,indicating that this method effectively plays the role of location information and can improve the accuracy of few shot image classification.Second,few shot learning method based on enhanced prototype feature representation.The pretraining-based meta-learning method solves the few shot problem effectively by pretraining the feature extractor and then fine-tuning it with the nearest center of mass meta-learning method.However,the results show that the improvements resulting from the fine-tuning steps are not significant.The reason is that in the pre-trained feature space,the base class has formed a compact cluster,while the new class is distributed with a large variance,which means that it is of little significance to fine-tune the feature extractor.During meta-learning,estimates of more representative prototypes are more valuable for few shot problems.Based on this fact,a meta-learning method based on enhanced prototype representation is proposed.Firstly,the original knowledge is introduced,and the representative attribute features are extracted as prior knowledge.Then,a prototype completion network is used to learn to complete the prototype with this prior knowledge.In order to avoid prototype completion errors caused by original knowledge noise or class differences,a Gauss based prototype fusion strategy is further used to combine the mean based prototype with the completed prototype by using unlabeled samples,so as to obtain a more representative prototype.Experiments show that this method can obtain more accurate prototype representation and improve the performance of few shot image classification.