Research On Few-shot Learning For Image Classification Based On Data Augmentation

Image classification for few-shot learning is a technology urgently needed for special scenarios,such as rare case images and industrial defect images,and is widely used in military,security and other fields.The fundamental problem of few-shot image classification is that there are few labeled samples,and the cost of manually labeling a large amount of unlabeled data is high.The data augmentation method can solve this problem and has been paid attention to and studied by a great deal of scholars.However,the few-shot image classification based on data augmentation has the problems of complex model,long inference time and poor distinguishability of samples generated by the augmented model currently.This paper studies form the perspective of model structure and loss function,and the few-shot image is augmented from the feature space.The main research content is as follows:(1)Aiming at the problem of complex data augmentation model for few-shot image classification,resulting in the long inference time,a simple tensor feature generator is proposed,which generates new tensor features by using the tensor features extracted by the backbone network directly,and augments few-shot images in the features space.Based on tensor feature generator,a rapid classification method for few-shot image is proposed,Tensor Feature-based Faster Classification Network(TFFCN).The TFFCN uses Res Net18 as the backbone network to extract features from the input images,and input the extracted image tensor features into the tensor feature generator to generate new tensor features,which can achieve the purpose of data augmentation.Using the augmented data to train the prototype classifier to obtain a satisfactory few-shot image classifier.Experimental results show that the classification performance of the TFFCN is better than the popular data augmentation methods for few-shot image classification,and the inference time is greatly reduced compared to the model before the improvement.When the backbone networks are Res Net18 and Res Net12,the inference time is almost reduced by up to49% and 24% with the increase of the number of generated tensor features,respectively.(2)Aiming at the problem of few-shot image classification based on data augmentation,the data generated has poor distinguishability,resulting in the classification accuracy needs to be improved.An effective method is proposed,Mixture Loss Function-based for Few-Shot Image Classification Network(MLFCN).In MLFCN,first,the improved adaptively spatial feature fusion(IASFF)is used to adaptively assign weights for different sample features that compute prototype tensor features,considering the different contributions of different sample features in the support set to the prototype tensor features.Second,add the inter-class mean squared error term into the mean squared error loss to obtain a mixture loss function(MLoss).It is used to train the tensor feature generator that extends the inter-class distance of the generated tensor features and makes the generated data more distinguishability.Experimental results show that the MLFCN has higher classification accuracy than popular data augmentation methods for few-shot image classification.