| Currently,there is the overfitting bottleneck in the recognition domain of deep learning.Effective training of a large number of parameters requires a sufficient number of labeled examples,meanwhile there is a large gap between the model’s ability to predict base classes and new classes.Inspired by the human ability to recognize new objects from a small number of reference examples,few-shot learning become a research topic,aiming to enable models to generalize to new classes with only a small amount of supervised data.Metric learning is widely adopted by models for few-shot learning due to its strong performance.In the metric architecture,the image features generated by the feature extractor have a significant impact on the prediction results.This paper thus optimizes the features during and after extraction respectively to improve the classification results.The main work is as follows:The distribution of the features is inferior to the optimal distribution,which induces the optimization idea of strengthening the good distribution and improving the poor one.Therefore,this paper proposes a simple but effective feature adaptation algorithm.After obtaining feature vectors,the prototype is firstly abstracted from the features of support set,and prototypes are adjusted by pre-classification and weighted summation,combined with the query set features.Secondly,the features of query set are used to adjust themselves by interactive selection and weighted summation.The algorithm allows the distribution in the feature space to converge to the ideal situation,thus improving the classification accuracy for few-shot learning.Recently,MLP has proved to achieve competitive performance to popular paradigms such as Transformer,and MLP-Mixer is a pure MLP tandem network that strikes a good balance between accuracy and computational resources,with the core idea of multi-level mixing of channel information and token information of feature maps.In this paper,it is improved to a parallel architecture and a fusion module is added to enable adaptive fusion of the information from token mixing and channel mixing,then it is introduced to few-shot learning i.e.to act as a backbone,aiming to optimize the feature extraction.Experimentally evaluated: The model with the feature adaptation algorithm achieves outperformance on several datasets such as mini Image Net and tiered Image Net compared to several baselines,and the MLP-Mixer possesses a stronger ability to represent novel images compared to common backbones.In addition,other experiments for hyperparameter,ablation,and cross-domain demonstrate the robustness and effectiveness of methods. |
