Few-shot Image Classification Based On Refined Features

As an important direction of machine learning,few-shot learning aims at solving machine learning problems where data are scarce,which are quite common in a real scenario.Recently,the rapid development of neural network has brought new ideas to the research of few-shot learning;In the area of computer vision,most researches for few-shot learning focuses on the problem of few-shot image classification.Many methods are proposed to solve this problem,including those based on data augmentation,those based on special design of model,and those based on different inference procedures.However,these methods inevitably have to perform feature extraction for each sample and each class involoved,to represent their semantic information.Such feature extraction process may be explicitly fulfulled by a part of the model,or not.Howsoever,this paper focuses on dealing with the potential problems caused by the lack of data during this feature extraction process.The specific work is as follows:1.Many of the existing few-shot image classification methods adopt a vector from the feature space as the feature of a certain class.These methods include many recent metric-based methods with a high performance.However,these methods tend to use the average or a linear combination of the features of labeled samples as the feature of the corresponding class,which is not quite robust to outlier samples.In order to solve this problem,the method proposed in this paper uses a large-scaled dataset to meta-train the class-level feature extraction module,by using the ideal class-level features of these large-scaled classes(due to the large amount of samples,the ideal state can be relatively easily approximated)as the additional supervision information of the class-level feature extraction module,in the hope that the class-level feature extraction module obtained by such meta-training can output class-level features close to the ideal state while only receiving information from one or a few samples of each class.The experimental results on the mini Image Net and tiered Image Net public datasets show that the classification accuracy of the proposed method can surpass other state-of-the-art methods by at most 12%.In addition,for the aforementioned ”ideal state of category features”,we also illustrates its calculation process and explains why it's feasibler both theoretically and experimentally in this paper.2.Transductive inference,i.e.,inference on the entire unlabeled set of samples instead of a single unlabeled sample,is a commonly used improvement for few-shot image classification methods when applying them to real scenarios.A widely used approach of introducing transductive inference into few-shot classification is to refine the features of samples based on extra information obtained from the distribution of the unlabeled set.When using the distribution of the unlabeled set to assist the classification,existing transductive few-shot image classification methods(i.e.,few-shot image classification methods that adopt transductive inference)only consider the predicted labels of the unlabeled samples,regardless of the reliability of these predicted labeles.However,the reliability(or,the uncertainty)of these predicted labels is also important,as the model does not always gives correct predictions in few-shot settings.As a result,we propose a transductive few-shot image classification method based on uncertainty in this paper.The method is built upon a widely used transductive few-shot classification structure,i.e.,using pseudolabels of unlabeled samples to update class-level features.Based on this structure,the method proposed in this paper further introduces uncertainty of the classification results of unlabeled samples when updating class-level features.The method uses the mutual information of the classification results to represent the uncertainty.Because it's hard to calculate the exact value of the mutual information,the method uses its approximated version,obtained by perturbing the classifying process with test-time data augmentation.The experimental results on four public data sets,including mini Image Net,tiered Image Net,Fewshot-CIFAR100 and CIFARFS,show that the classification accuracy of this method is higher than that of most state-of-the-art transductive and non-transductive few-shot classification methods by 1.5%-17.7%.This paper also visualizes the class-level features generated by other state-of-the-arts and the method proposed in this paper,to qualitatively illustrate the improvement brought by introducing uncertainty to transductive few-shot classification methods.