Small Sample Image Classification Based On Deep Learning

Image data increases explosively along with the advancement of Internet technology and the improvement of information level.How to select the image data of their own demand In all kinds of image databases has posed certain challenges to image classification technology.recently with fast development in the field of deep learning,manual annotation characteristics of image classification has gradually given away with deep learning image classification technique which has high accuracy and recognition efficiency.in the course of training of deep learning, however,there exists some deficienciesparameters is difficult to adjust,the training sample dem and is bigger and the training time long shortcomings and etc.To deal with those problems,it is very meaningful to carry out researches with a small number of sample cases about how to efficiently use deep learning for image recognition.in addition,the adaptability of the deep learning model studied under the condition of different samples would be enhanced.Many problems are identified for small sample recognition,for example,easily overfitting and poor generalization.In this paper,we put forward classification and recognition model based on deep learning,the optimization direction is mainly based on the following two aspects.The first is the enhancement of the number of image samples,and the second is the recognition model optimization.When the sample size is increased,we put forward the technology of combining generative model with image preprocessing.First,we use the fully connected generation model to sample enhancement.Aiming at the problem of too many parameters in the fully connected neural network,the convolutional neural network is used instead of the fully connected neural network for image training.Due to the randomness of the generated sample images,the conditional generation model is used to generate samples.The sample set contains labels,which can be used well in the follow-up supervised classification learning.In order to produce fuzzy samples in conditional generation model,an image edge detection technology based on wavelet transform and adaptive mathematical morphology is proposed.This method can overcome the problem of edge blur.In addition,because of the noise in the generated samples,this paper proposes an image denoising technique based on the combination of empirical mode decomposition and sparse representation.The optimized denoising model can preserve the edge details of the image while removing the noise.Therefore,using the fusion model to enhance the number of image samples can expand the sample to the next part of the classification model.In the small sample classification model,migration learning is able to achieve higher recognition efficiency in fewer training samples.Transfer learning can achieve higher recognition efficiency in fewer training samples.In this paper,combining the generative model with the transfer learning,we use the fusion model to recognize small samples.The Inpection-V3 model is used to train the migration learning.Compared to the simple convolution neural network,the migration learning can improve the generalization ability of the model,and the training efficiency of the model has also been greatly improved.In the absence of samples,the accuracy of recognition based on transfer learning is greatly improved when samples are identified.when the sample is missing,in this paper,the model is merged and classified by increasing the combination of imitation samples and transfer learning.Comparing the standard data set and the shoot leaf image,the algorithm is better than the simple use of the migration learning and convolution neural network in the accuracy rate.It has a good reference meaning for the deep learning classification in the experimental environment with insufficient sample size.