Research And Implementation Of Small Sample Image Classification Algorithm Based On Metric Learning And Data Enhancemen

Image categorization has been used in various contexts as a significant piece of work.At present,the mainstream image classification algorithms are based on deep learning,and most deep learning networks can have good performance.It not only requires excellent models,but also requires a large amount of labeled data.But in actual applications,a lot of things are hard to get and have large costs.In order to solve the problems of appeal,this paper aims to improve the accuracy of small sample image classification by combining metric learning and data augmentation,and apply the model to the field of blood cell classification,to a certain extent,to ease the pressure of obtaining large-scale supervised data sets.The main research results and innovations are as follows:(1)Aiming at the problems of insufficient feature extraction and lack of nonlinear relationship measurement between features in the traditional metric learning model,this paper combines the attention mechanism and residual structure into the metric learning feature extractor,so as to enhance the extraction performance of network features and alleviate the overfitting pressure of the network.For the metric classifier,the metric learning classifier in this paper uses a convolutional network that can learn nonlinear relationships to improve the measurement ability of the network.Finally,compared with the current mainstream small sample classification model,the improved metric learning model(SRRN)has a better effect on the classification of the mini Image Net dataset,and it is verified on the blood cell dataset that the improved network model still has a certain generalization ability.(2)In order to solve the over-fitting problem caused by the lack of samples,this paper proposes an improved generative adversarial network(ALGAN)to improve the quality of images and add it to a small number of sample classifications to improve classification accuracy.Specifically,this paper uses self-attention mechanism to find a good balance between improving receptive field and reducing the number of parameters.adding label information in the generator,from unsupervised generation to supervised generation;then the batch normalization in the discriminator is replaced by spectral norm normalization and the wassertein distance is used to redesign the loss function to stabilize the training process of the model.By comparing the images of CIFAR-10 and MNIST,the results show that the image quality generated by ALGAN is better than that of conventional GAN model.A set of mini Image Net small samples are generated by ALGAN,which can enhance the data of small sample images,t so resolving the over-fitting issue with small sample learning and somewhat increasing the model’s classification accuracy.