Research On Few-shot Learning Method Based On Deep Feature

Recently years,deep learning has achieved good results in many computer vision tasks.But deep learning method requires a large-scale dataset to train the model,and the model convergence is slow.In contrast,the human visual system is able to recognize new objects after only observing one or a few instances.This significant gap between the human visual system and deep learning models has aroused research interest in few-shot learning.The purpose of few-shot learning is to build a classifier that have a good classification performance in new classes where each class has only a few numbers of labeled samples.At present,there are many excellent solutions in the field of few-shot learning,where the metric learning method is widely used because of its simplicity and efficiency.It firstly learns the feature of the example through the feature network,and then directly uses the nearest neighbor in the feature space.Based on the metric learning method,this paper further proposes an improved solution of the feature network and the metric manner,which can efficiently process few-shot learning tasks and obtain the best results on this task.The main research contents are as follows:1.Explain the basic form and related methods of few-shot learning.The problem definition and task form are systematically introduced for few-shot learning,and then the related methods of few-shot learning are introduced in detail.Including the data augmentation methods,the meta-learning methods and the metric learning methods.Among them,meta-learning methods are divided into memory-based methods and optimization-based methods in detail.This paper introduces the ideas of various methods in detail and analyzes their advantages and disadvantages.2.This paper based on metric learning and inspired by the idea of prototypical network,a label feature network is proposed.Single-scale image features ignore the detailed information under different scale images and the average class prototype are not well evaluated the different contributions of each support set example in the class to the class prototype.The concept of multi-scale features and label features are proposed.The label feature network includes two modules: a feature module and a label feature module.The feature module is used to learn image features,and the label feature module is used to learn the prototype expression of each class.First,the multiple scale images are obtained by processing the image,the feature module takes the multiple scale images as input to learns the multi-scale features of the image,and then the label feature module takes the depth concatenation of features in the class as input to learns the label feature of each class.3.The few-shot learning methods based on metric learning learns the features of the images,then calculates the similarity between the features and the classification is performed by looking for the nearest neighbor.Therefore,how to effectively calculate the similarity between features is the difficulty in few-shot learning.Currently,the similarity calculation mainly uses predefined distance metric.The predefined distance metric based on the given learning features,which depends on the quality of the learned features.When the distinction information of features learned by feature network is not sufficient,the predefined distance metric often limited.Due to the similarity measure depends on the learned feature quality,a metric agnostic method is proposed,which use a neural network to calculate the matching degree between features.Similarity calculating jointly learning with feature extraction can improves the classification effect.