Research On Meta Learning In Few-Shot Image Classification

Deep learning can be used to solve many practical problems and make people’s lives smarter and more convenient.In practical applications,deep learning has a large amount of training data,describes the problem with a large number of parameters,and can fit the problem well.However,when the data is insufficient,deep learning often encounters the challenge of overfitting or generalization.When the model is overfitted,the generalization ability of the model is very low.Few-shot learning is designed to solve machine learning tasks in such data-limited situations.Meta-learning is an idea that effectively implements few-shot learning.The goal of meta-learning research is to achieve model transfer between more different tasks,and to adapt to more different tasks by decoupling model fitting and model generalization.At the same time,the research goal of meta-learning is to achieve general artificial intelligence,so that machines can learn to update themselves and adapt to tasks with lower similarity through interaction with other machines and the environment.This topic takes small sample image classification as the research object,studies how to better use meta-learning to solve problems encountered in practical applications,and explores how to further optimize the current meta-learning algorithm,improve the accuracy and generalization of the algorithm,Accelerated meta-learning algorithms land in more application scenarios.The main work is as follows:(1)The MAML algorithm based on Bayesian improvement is designed and applied to the actual industrial defect image classification scene.Industrial defect classification is a typical few-shot learning scenario.The algorithm combines Bayesian decision theory with Bayesian variational inference for labeling prediction,and considers the posterior distribution of the initial parameter values in the task training set.Theoretical analysis and experimental results show that the algorithm proposed in this thesis can effectively improve the accuracy of the MAML algorithm,and can be used as a solution for industrial defect classification.(2)A batch training strategy is introduced for the parameter optimizer.Experiments show that,compared with the previous sequential optimization strategy,this strategy can effectively suppress the oscillation of the model accuracy curve during the training process and improve the robustness of the model..(3)A meta-learning algorithm based on task attention mechanism is designed,and the task attention mechanism is extended to the meta-learning strategy based on model initialization and parameter optimization.Multiple experiments on different datasets show that adding a task attention mechanism helps the meta-learning model to obtain better generalization ability.