Research On Few Shot Image Classification Based On Meta-learning

In the image researches,algorithms based on deep learning have surpassed the recognition level of humans in classification tasks.However,when the training samples are scarce,the image classification model based on deep learning often causes overfiting and poor performance.Then it fails to achieve the expected classification effect.In order to solve this problem,few shot learning has become a current research hotspot.Few shot learning can quickly learn key knowledge from a small number of data samples to be competent for new tasks.In addition to solving deep learning problems in data sparse scenarios,it is also of great significance for exploring artificial intelligence technology.But in the face of data type drift,the performance of the model will be greatly reduced.In order to address the above problems,this thises proposes a few shot image ensemble classification algorithm based on meta-learning,which can reduce the difficulty of training a single classifier by training multiple weak classifiers on the feature channel.At the same time,the meta-learner is used to generate the parameters of each weak classifier according to the channel characteristics and reduce the performance loss of the model when the data type drifts.In addition,to enhance the classification performance of the model on fine-grained data sets,a multi-granular feature fusion few shot learning method based on meta-learning is proposed.Taking into account the difference of the feature information of different granularity layers,the information of different feature layers is fused in the feature extraction to improve the accuracy of the model's recognition of fine-grained feature data.The specific research content is as follows:1.With the help of the attention mechanism,according to the importance of different channel features to the task,multiple weak classifiers are designed.And the similarity between samples is calculated on multiple channels at the same time.Then the result of the weak integrated classifier is used as the classification result of the model.The integrated classifier simplifies the difficulty of training a single classifier and improves the final classification performance of the model;2.In the defined similarity measurement network,the multi-layer perceptron is used to replace the measurement formula.The model directly generates the multi-layer perceptron parameters according to the current task.Compared with the use of Euclidean distance,cosine similarity and other measurement formulas,the end-to-end design reduces the difficulty of training the feature extraction network,and improves the generalization ability and classification accuracy of the model.Finally,the problem of severe model performance degradation has been effectively solved while the data type drifts;3.Based on the multi-granularity cognition,the feature vectors of different granularities are merged in the process of model feature extraction to improve the model feature extraction ability.At the same time,the graph convolutional neural network is used to calculate the similarity between samples.Thereby,the classification accuracy of the model has been improved on the fine-grained data set.