A Research Of Few-shot Classification Based On Meta-learning

Few-shot learning is designed to learn about object categories from a single or a small number of training samples,which is critical to deep learning based on large amounts of data.Deep learning can solve the problem of few-shot learning by meta-learning using the previous experience to learn how to learn.Therefore,this thesis takes the problem of few-shot classification as the research object and studies how deep learning uses meta-learning to quickly learn and generalize from a small number of samples.The main contents are as follows:Firstly,conducting metric learning on a wide range of tasks allows the deep learning method to make good use of previous empirical knowledge,but it is subject to the quality of feature extraction and the selection of metrics for support and target sets.Therefore,in view of the above problems,this thesis designs a multi-scale relation network.The output of the third and fourth modules of the convolutional neural network consisting of four convolutional modules is spliced in the depth direction to obtain multi-scale features of the support set and the target set.Multi-scale features of k samples of each category in the support set are averaged,then combined with the multi-scale features in the target set by the way of absolute values of feature subtraction.The results on the benchmark sets show that the multi-scale relation network is simple and effective,which not only alleviates the over-fitting situation,but also improves the classification accuracy.However,to ensure that the metric apply to all tasks,the few-shot learning based on metric learning must ensure the homologous distribution of the task set.Secondly,considering that the methods of few-shot learning based on metric learning must guarantee the homologous distribution of the task set,this thesis adds the model-agnostic meta-learning algorithm to the multi-scale relation network,then designs the multi-scale meta-relation network.The multi-scale meta-relation network adopts the idea of meta stochastic gradient descent.The inner learning rate is taken as learning vector and learned together with model parameters to further improve the performance of the learner.In the meta-training process,the model-agnostic meta-learning algorithm is used to find the optimal parameters of the model,but in the meta-validating and meta-testing process,the inner gradient iteration is cancelled.The experimental results on the benchmark sets show that the multi-scale meta-relation network enables the learned metrics to have a stronger generalization ability,which not only improves the classification accuracy on the benchmark sets,but also avoids fine-tune when using the model-agnostic meta-learning algorithm.Finally,the multi-scale meta-relation network is introduced into the field of fine-grained and remote sensing image recognition.The results show that the multi-scale meta-relation network can be applied to the fine-grained and remote sensing image classification under small sample conditions.