Research Of Few-shot Learning Algorithms Based On Metric Learning

In recent years,deep learning has played an important role in many fields and achieved great success with the help of large-scale data.However,there are still some drawbacks.For example,in order to achieve better performance in specific tasks,deep learning models often require a long time of training with large amounts of data.While actually it is expensive to collect and label large-scale data.Therefore,few-shot learning has gradually become an important research topic in computer vision classification tasks.Few-shot learning presents a challenge that a classifier must quickly adapt to new classes,given only a few labeled examples of each new class.Based on the situation described above,in this paper,we focus on the study of metric-based methods in few-shot learning.Our main contributions are as follows:By analyzing the problem that convolutional layers in relation networks can only compare local features,we introduce deformable convolutional neural networks as the feature extractor in relation networks,which makes the extracted features have certain global characteristics,and alleviates the problem that relation networks can only compute the relationships of local regions and get an inaccurate result.As for the metric module in relation networks,we also propose a dual correlation attention mechanism that can capture the global information of inputs to densely aggregate into each spatial position of outputs.In this way,despite of the local connectivity,the subsequent convolutional layer can involve the global information when computing relationships for local regions,and adaptively compare related fine-grained features.Experimental results show that by introducing only a small number of parameters,the dual correlation attention mechanism can significantly improve the metric ability of relation networks and achieve better performance.Finally,we propose an effective structure of metric model that contains parameters,which is based on the proposed multi-head adaptive factorization machines.By referring to the idea of the factorization machines,we first convert the origin parameters of the weighted metric model into the form of factorization parameters.Then we use fully convolutional neural networks to generate corresponding factorization parameters adaptively for each pair of input images.After that,we ensemble several adaptive factorization machines as the final proposed model.Experiments show that the proposed model can acquire a better metric ability than other metric modules without paremeters,and achieve better performance.