Research On Few-Shot Image Intelligent Classification Method Based On Meta-Learning

Deep learning has achieved good results in many fields,but its success depends on the amount of available data and the ability of computing resources.Due to the difficulty of obtaining effective annotation data and the high cost of annotation,the traditional deep learning method is not universal for image classification tasks in some specific cases,which greatly limits the research and application of deep learning in the case of few-shot.In recent years,meta-learning technology has become a research hotspot.It improves the accuracy of image classification in the case of few-shot,but there are also some problems,such as data imbalance,the poor performance of the feature extraction network,and the inability of the feature extraction network to learn high-level semantic information and low-level feature information at the same time.Because of the above problems,this paper mainly focuses on using meta-learning to deal with image classification in the case of few-shot.The specific research contents and results are as follows:Firstly,aiming at the data imbalance problem of few-shot learning in the real world,the MAML algorithm based on meta-learning is improved and optimized.The data set encoder and balance variable generator are designed to learn statistical information,and two balance variables are generated through statistical information for the model optimization part to adjust the learning rate of the model for a category to reduce the impact of data imbalance.Secondly,aiming at the problem of poor performance of feature extraction network,convolution kernel is dynamically generated for each input,and adaptive convolution operation is carried out to improve the ability of feature extraction.In addition,embedded propagation is used as a regularizer for manifold smoothing,which is placed at the top of the feature extraction network to expand the range of decision boundaries and enhance the information representation of categories.Experiments show that on mini Imagenet and tired Imagenet data sets,under the conditions of 1-shot and 5-shot,when the feature extraction network is conv-4,the classification accuracy is improved by 2.04%,3.20%,4.61%,and 3.61% respectively compared with EPNet.Thirdly,aiming at the problem that the feature extraction network can not learn high-level semantic information and low-level feature information at the same time in the meta-learning method based on measurement,this paper designs a multi-scale feature weighted fusion extraction network based on the relational network.Through the improved multi-scale weighted feature fusion module,the feature maps with different sizes and levels are fused to enhance the ability of network feature extraction.Finally,given the lack of purposeless learning in most cases of feature extraction networks,channel attention mechanism,and spatial attention mechanism are introduced into the top part of the feature extraction network to impose attention weight on channel level and pixel-level features,to quickly screen high-value information from a large amount of information with limited attention resources.The experiments show that the method proposed in this paper improves the accuracy of image classification and enhances the feature extraction ability of the model to a certain extent.The effectiveness of each functional module is verified by the ablation experiment.The experiments show that on mini Imagenet and tired Imagenet data sets,under the condition of 1-shot and 5-shot,the classification accuracy is further improved by 0.47%,1.08%,1.71% and 0.60% respectively compared with EPCINet.