Semi-supervised Few-shot Classification Method Based On Meta-learning

The deep neural network model has achieved good results in the task of image classification,but this is based on a large number of labeled samples.If the current label samples are small,it is easy to produce over-fitting phenomenon,and the model’s generalization ability cannot be effectively guaranteed.Different from traditional supervised learning,meta-learning corresponds to the task distribution composed of multiple tasks,and adapts to small-sample scenarios by learning each task,which makes it possible for deep neural network models to be used in small-sample learning.Enlarging the dataset by introducing unlabeled data is an effective solution for meta-learning methods to solve the problem of few-shot,and has gradually been paid attention to by researchers.However,there are still two problems: how to generate pseudo-labels for unlabeled data by using a small number of samples and how to make full use of the generated pseudo-labeled samples.Aiming at the above two problems,this paper proposes a semi-supervised few-shot classification algorithm based on gradient re-optimization and a semi-supervised few-shot classification algorithm based on self-training,and implements a semi-supervised few-shot classification system based on meta-learning.The specific work includes three aspects:(1)A semi-supervised few-shot classification algorithm based on gradient re-optimization is proposed.On the one hand,a pseudo-label generation algorithm based on self-consistent regularization is proposed to generate pseudo-labels for unlabeled data in the case of few-shot.On the other hand,through the operation of gradient re-optimization,a network model that is more suitable for the current task is obtained to achieve a better detection effect.At the same time,the first-order approximation of the meta-task is used to replace the second-order guide information,which effectively reduces the time complexity of the algorithm.(2)In order to make full use of the intrinsic data characteristics of the same unlabeled sample after enhancement,this paper further proposes a semi-supervised few-shot classification method based on self-training.When processing unlabeled data,the data enhancement is divided into weak enhancement and strong enhancement,and a threshold is set.The corresponding category of the pseudo-label is judged by comparing the relationship between the predicted value and the threshold,and then the network model is updated in a self-training manner.Experimental results show that this method has a better classification effect than the existing few-shot classification methods.(3)Based on the semi-supervised few-shot classification algorithm based on self-training,this paper designs and develops a semi-supervised few-shot classification system based on meta-learning,which visually displays the results of the algorithm’s classification of few-shot,so as to facilitate the evaluation of the effectiveness of the model.The test results show that the system not only has fast processing speed,but also has good human-computer interaction.