Deep Learning Image Classification Research Based On Few-Shot And Unbalanced Data

Image has become an important form of expression for a large amount of information and an important source for people to obtain information because of its advantages such as rich forms of expression and intuitive reflection of things.With the advent of the information age,more attention has been paid to the image classification techniques that keep up with the times in order to extract deeper and more important contents from a large number of high-dimensional images.Due to the rapid development of deep learning in recent years,deep learning-based image classification algorithms have gradually replaced traditional image classification algorithms with their superior performance.However,such approaches usually require large and balanced labeled data to train the network to achieve better classification performance.Insufficient or unbalanced training data greatly affects the model performance,which in turn limits the practical application of these methods.Therefore,few-shot learning and unbalanced learning have become essential research hotspots in the field of image classification nowadays.In order to solve the problem of image classification under the background of small sample and unbalanced data,this paper starts from deep learning method and focuses on two typical problems of synthetic aperture radar(SAR)image classification under small sample data and breast cancer histopathological image classification under unbalanced data respectively.(1)Aiming at the problem that traditional convolutional neural network(CNN)model square convolution kernel has a large amount of calculation,poor effect on fine feature extraction and the classification performance degrades with the increase of the network depth in the SAR image classification work under small sample data,we propose a convolutional neural network based on asymmetric parallel convolution and residual learning(APCRLNet).The proposed asymmetric parallel convolution structure can extract the width and height features of SAR images,which enhances the generality of the model.Meanwhile,a residual learning method based on skip-layer connections is applied to this problem to extract the deep-level features of images and better train the deep network model.The model improves the classification performance while reducing the computational complexity.Experimental results on the moving and stationary target acquisition and recognition(MSTAR)dataset show that APCRLNet achieves higher classification accuracy with less training time compared with Basci-Net and RL-Net.APCRLNet has better classification performance under full training samples or part training samples.(2)Aiming at the problem that the loss function of the original capsule network model only focuses on the overall information of the SAR image and ignores the information differences between different classes,and the lack of high-level semantic information leads to poor model classification performance in the SAR image classification work under small sample data,we propose a multi-dimensional parallel capsule network with class separable loss(MdpCaps-Csl).The proposed class separation loss function based on angle cosine similarity improves the feature extraction ability of the network for SAR images by reducing intra-class differences and increasing inter-class differences in the process of feature extraction.Meanwhile,the multidimensional parallel capsule module takes the feature maps obtained from different levels of convolution as input respectively,and performs capsule encoding on different feature maps,which can improve the network classification performance.The experimental results on the MSTAR dataset show that MdpCaps-Csl outperforms most traditional machine learning methods and deep learning-based methods,and the fewer training samples are used,the more obvious the results of the comparative experiments are.In the case of a very small number of training samples,the classification performance of MdpCaps-Csl can still achieve a high level even when training data with different rotation angles are lacking.(3)Aiming at the problem that the bilateral branch network(BBN)model is not effective in extracting fine spatial features and the lack of low-level structural features which leads to poor model classification performance,and it requires artificial preset weighting parameters through experience which causes large subjectivity and limitations in breast cancer histopathology image classification work under unbalanced data,we propose a meta-adaptive-weighting-based bilateral multi-dimensional refined space feature attention network(MAW-BMRSFAN).The method consists of the classification model BMRSFAN and the parametric ? learning model MAWN(meta adaptive weighting network).The proposed refined space feature attention module(RSFAM)based on convolutional long short term memories(ConvLSTMs)model is designed for extracting finer lesion information in breast cancer histopathology images to improve feature learning capability of unbalanced image classification tasks.Meanwhile,MAWN can be used to model the mapping relationship from balanced meta-dataset to unbalanced dataset,making it more flexible to find appropriate weighting parameter for BMRSFAN.The experimental results on the BreaKHis dataset show that for different imbalance coefficients,the classification performance of MAW-BMRSFAN is much better than most of the methods proposed so far.Moreover,the MAW-BMRSFAN also achieves ideal classification performance in extreme unbalanced situations.In summary,this paper develops a series of studies on image classification methods in the context of small sample and unbalanced data from deep learning-based image classification techniques with the goal of achieving better results in practical applications,and improve classification performance and generalization ability on different research problems.