Research On Deep Learning Methods For Small-sample Image Classification

Image classification of a small number of training samples is a difficult problem in the field of artificial intelligence and computer vision.In existing research,deep learning models have achieved the most advanced performance in visual tasks(such as image classification).However,the generalization performance of the model under a small number of training samples is not satisfactory.At present,many researches have made use of domain adaptation and data enhancement methods to make up for the lack of sample data.Regularization techniques,ensemble learning,and the method of changing the distance between classes are used to alleviated the model 's overfitting and poor generalization performance.However,these methods still can not meet the requirements of the model classification ability of the image classification of a small number of training samples.Based on the deep convolutional neural network,this paper has completed the following three tasks for the above problems.First,a cross-entropy loss function method for image classification of a small number of training samples is proposed.Existing cross entropy loss functions only focus on the probability that samples are assigned to the correct class,but not the probability that samples are assigned to the wrong class.For the problem of multi-class cross entropy loss function,the cross entropy loss function is improved under the condition of a small number of training samples,so that the probability that the sample is classified into the correct class is considered,while the probability that the sample is classified into the wrong class is considered,thereby further improving the generalization ability of the model and improving the stability of the model.This work is experimentally verified in the Stanford Cars-196 dataset,UIUC-Sports dataset,LabelMe dataset and CIFAR-10 dataset.Relevant experiments show that the proposed method has better advantages than the existing methods.Second,an ensemble learning approach for classification of small-samples is proposed.It is a problem to be solved that deep neural networks are easy to overfit on a small number of training sample datasets.In the existing research methods,ensemble learning is a good method to overcome model overfitting and reduce model variance.However,due to the large randomness of the deep neural network,the existing ensemble learning method based on the deep neural network still has the overfitting problem on the small-sample image data.Therefore,this paper proposes a new ensemble network based on the methods of boundary learning and prototype learning.This work is experimentally verified in LabelMe dataset and Caltech101 dataset.Experimental results show that the proposed integration method can achieve better generalization performance than the existing methods.Third,a method of extracting discriminative features for small-sample classification is proposed for small image sample.However,many studies have shown that even if the features within the class are close and the distance between classes is long,the performance of the model is not substantially improved.In order to solve this problem,a new neural network based on prototype learning and relational network is proposed in this paper.This work is verified in UIUC-Sports dataset,LabelMe dataset,15 Scenes dataset and BMW-10 dataset.Experiments on four datasets show that the proposed method has better performance than the existing methods.