Research On Structure Of Few-shot Image Classification Neural Network Based On Metric Learning

Under the current environment of rising artificial intelligence and machine learning,deep learning model has made a breakthrough in the field of image classification based on large samples.However,in practical applications in real life,the amount of data that can be provided is not enough.In order to solve the problem of overfitting or underfitting of network model caused by insufficient data,data enhancement,self-supervision and semantic information using methods are introduced successively,and meta-learning methods based on transfer learning,optimization and measurement are introduced to improve the classification accuracy of the model.However,the existing Few-shot image classification methods often adopt a single form of features,and do not pay attention to the discriminative local information of image features.Therefore,in order to improve the classification accuracy of the model under different small number of sample data sets,improve the generalization ability and robustness of the model,this paper completed the following three works according to the existing metrics-based Few shot image classification method:1.A robust dual prototype network for Few-shot image classification is proposed.At present,most image classification methods only adopt one featu re form and use a single metric to measure the distance between samples,which often results in poor model accuracy and generalization.Therefore,in order to improve the robustness and classification accuracy of classical prototype networks,a new robust dual prototype network(DPN)is proposed to solve the problem that the model measurement form is too simple.The network not only pays attention to the distance of the original feature,but also adds disturbance noise to the image and calculates the distan ce of the noise feature.Finally,the model is predicted under two different indicators,so that the model can learn more representative and robust class prototype,so as to obtain better generalization performance.2.A local content extraction network fo r Few-shot image classification is proposed.In view of the fact that most of the current Few shot image classification methods only extract the basic features of the image for category prediction,lack of consideration of the local features of the detailed semantic information.Therefore,to solve the problem of inaccurate feature extraction from the model,a local content extraction network(LCEN)is proposed.The core module of the network is a plug and play local content extraction module(LCEM),which is used to learn the discriminative local information of the image,so that the model is more focused on extracting local features of the image and filtering redundant information.Most of the background is removed,and more accurate information is extract ed compared with the original basic features to assist the classification of the Few shot images.3.A feature cross reconstruction network for Few-shot image classification is proposed.The feature reconstruction of an image can preserve more details of t he appearance,which helps to classify the image into the category with the least reconstruction error.In order to solve the classification problem of Few-shot images based on feature reconstruction,a feature cross reconstruction network(FCRN)was proposed.The network extracted basic features by embedding modules,the local content extraction module learned local features,and the cross reconstruction module(CRM)cross-fused and reconstructed basic features and local features by ridge regression formula.The classification score is calculated based on the weight of the reconstruction error of the cross reconstruction task,and the corresponding weight is learned from the training process,so as to obtain a higher classification accuracy.In this paper,the background and status quo of the current Few shot image classification are analyzed,and three network models or frameworks based on metric learning are proposed.The improved methods proposed achieve the optimal performance of the current Few shot ima ge classification,and improve the generalization ability of the model to different data sets.