The Research On Small Sample Image Classification Method Based On Deep Learning

With the rapid development of deep learning,image classification based on big data has made important progress,and it has surpassed the level of human recognition in many datasets.However,there is still much room for improvement in small sample image classification tasks.Given that the number of things in the real world is consistent with the long tail distribution,that is,the number of most things is very small,and human learning a concept does not require thousands of training data,small sample image classification has become a research hotspot in the field of computer vision and machine learning.This paper is mainly based on deep learning,and studies image classification methods under a small number of training samples.The main work includes the following aspects:(1)A neural network loss function for small sample image classification is proposed.Although neural networks based on cross-entropy loss functions have achieved great success in many fields,when there is less data,it is difficult to avoid over-fitting problems.To solve this problem,this paper proposes a new neural network loss function.In the model training process,the proposed loss function adaptively adjusts the distance between decision boundaries,so that the model extracts more discriminative features.Experimental results on three small sample datasets show that the proposed method has better generalization ability than the existing regularization methods.(2)An orthogonal classification layer for small sample image classification is proposed.Deep neural networks usually model a huge function space,when the training data is small,it will encounter severe over-fitting problems and increase the instability of the model.In order to solve this problem,this paper proposes an orthogonal classification layer,which removes the partial connections in the traditional fully connected layer,so that the weight vectors between different classes remain orthogonal during the training and test process.The experimental results on four small sample image classification datasets show that the proposed method can effectively alleviate the over-fitting problem and increase the stability of the model.(3)An interactive neural network for small sample image classification is proposed.Deep neural networks usually contain a large number of parameters and require a lot of training data.When the training data is small,the neural network containing a large number of parameters is easy to overfit the training data,and there are problems such as large variance of the model and poor generalization ability.To solve this problem,we propose a new ensemble method for small sample image classification,called InterBoost.During the training process,InterBoost first randomly generates two complementary datasets,respectively training two base classifiers of the same structure,and then generates two complementary data sets for further training through the interaction between the two base classifiers trained.This interactive training process will iterate until the stop condition is met.In the test phase,combine the output of the two networks to get a final prediction.Experimental results on four small sample data sets show that the integration method is superior to the existing integrated classification method.(4)A snapshot ensemble method for selecting model initialization is proposed for small sample image classification.The snapshot ensemble method obtains multiple different base classifiers by searching for multiple local minimums of the loss function during the training process,and performs integrated prediction during the test phase.However,when the training data is small,the initialization parameters of the model will greatly affect its classification performance,resulting in instability.To solve this problem,we propose a new ensemble method called selective snapshot ensemble learning method.During the training process,poor model initialization is monitored and discarded by setting thresholds.The results on two small sample classification datasets show that this ensemble method is superior to the snapshot ensemble method and other ensemble methods.