Research On The Method Of Few-shot Deep Learning For Image Recognition

The key idea of deep-learning based few-shot learning is that hoping deep neural network can be pretrained on a large number of labeled base category examples,so that this network can realize the rapid generalization recognition of novel categories with a few labeled examples.One of the main methods to realize few-shot learning is meta-learning.However,meta-learning based method not only need lots of training time and computing resource,but the performance is not satisfactory.Recently,classical supervised representation learning methods without meta-learning have been proved to achieve competitive results compared with metalearning based methods for few-shot learning,which shows amazing potential.In this thesis,to further improve the performance of few-shot learning models,based on the classical supervised representation learning schema,we propose two improved few-shot learning methods.To sum up,the main research work in this thesis can be listed as follows:Firstly,in order to learn more powerful visual feature representation,we proposed a novel mixed feature representation learning(MFRL)for few-shot learning,the loss function of this model not only includes a classical supervised softmax cross-entropy classification loss,but also an extra unsupervised image transformation parameters reconstruction loss,which makes the model can jointly learning the visual feature representation specialized to the category labels and the category-agnostic transformation-equivariant visual representation,so as to improve the few-shot learning performance.Secondly,in order to improve the discrimination ability of visual classifier,we proposed a knowledge transfer network(KTN)for few-shot learning.By introducing the semantic knowledge information contained in category labels,the proposed KTN model jointly incorporates visual feature learning,semantic-vision knowledge inferring and classifier learning into one unified framework,which makes the more discriminant visual classifier for novel categories can be generated adaptively,thus significantly improving the performance of few-shot learning.