Image Classification Based On Meta-Learning

Deep learning algorithms have achieved great success in many fields such as image classification.However,the superiority of these algorithms relies on large-scale datadriven training processes,which limits the scalability of deep learning algorithms.Therefore,the study of few-shot learning(FSL)algorithms can reduce the dependence of neural network models on labeled examples.This paper has carried research from training strategies,metric learning,module design,based on the meta-learning paradigm,The main work is as follows:(1)Inspired by human education,we propose a meta-learning training framework based on curriculum learning to guide the model to learn tasks from easy to difficult.The core idea of the method is to build a sequence of observable class set following the training progress before episodic sampling.Each class subset tends to select elements that are separated each other.Meanwhile,the capacity of the class set is increasing gradually.Experimental results demonstrate that the curriculum-based training framework can be integrated with a variety of meta-learning algorithms,improving the final classification performance of the model.(2)We propose a dual-branch transductive FSL method based on the optimal transport theory.The method projects the query sample and the class center into the cosine space,then optimizes the transport matrix to maximize the similarity between these query features and the class centers.Util the transport matrix and the cosine similarity between the query feature and the class center to build a dual-branch prediction.Two prediction probabilities are fused using linear interpolation.Experimental results demonstrate that the proposed method can achieve competitive results on multiple FSL benchmarks.(3)The FSL classification task cannot meet the independent and identical distribution(i.i.d)assumption of batch normalization(BN).We introduce cross-tasks statistics information based on existing normalization technology to improve the stability of BN’s moment.However,our method increases the number of model parameters,when we use it,we need to balance the benefit of the stable moment estimation and the risk of overfitting.