Research And Implementation Of Few-shot Recognition Algorithm Based On Metric Learning And Data Augmentation

In recent years,with the rapid development of deep learning,few-shot learning has gradually become a current research hotspot with excellent scenarios and application potential in the fields of medical treatment,rare species and so on.Few-shot recognition aims to learn information about object classes from one,or only few labelled images.The biggest challenge is that most of the current few-shot recognition models have low accuracy,and there is still much room for improvement.In order to effectively improve the accuracy of few-shot recognition,the idea of meta learning is used to research the few-shot recognition algorithm based on metric learning and data augmentation in this paper.The main work is summrized as follows:(1)Aiming at the problems of the insufficient feature extraction and the difficulty of fully fusing features in the metric function for the current metric learning-based few-shot recognition models,an end-to-end,metric learning-based model is proposed in this paper,called multi-scale decision network based on feature fusion and feature weighting(MSDN).First,the feature extraction network is used to extract the multi-scale features of the image,which can effectively exploit the feature information of each layer.Then,the relation network is used to improve the way of feature concatenation,which introduces feature fusion and feature weighting to measure the feature similarity of the support set and query set.Finally,MSDN achieves the state-of-the-art accuracy result on Omniglot and minilmageNet datasets compared with popular few-shot recognition models.(2)Aiming at the problem of over fitting caused by the lack of data in the few-shot recognition model,this paper uses generative adversarial network(GAN)to increase training examples for few-shot recognition.At present,the traditional convolutional GAN treats spatial and channel-wise features equally,which causes the features extracted by the convolutional network to fail to capture key information.To address the issue,a generative adversarial network based on convolutional block attention module(CBAM-GAN)is proposed in this paper.CBAM-GAN adds the convolutional block attention module after some convolution operators to adaptively rescale spatial and channel-wise features,which can enhance salient regions and extract more feature details.The comparative experiments of image generation on the MNIST and CIFAR-10 datasets show that compared with the traditional convolutional GAN,CBAM-GAN can effectively improve the quality of generated images.Finally,CBAM-GAN generates a batch of miniImageNet images to expand the few-shot dataset and slightly improves the accuracy of few-shot recognition.(3)In the work of using CBAM-GAN for data enhancement,the accuracy of few-shot recognition is only slightly improved.This is mainly because that using CBAM-GAN will generate a small number of noisy images,which causes the loss of recognition gain.To address the issue,a progressive data generation network based on CBAM-GAN(Progressive CBAM-GAN)is proposed in this paper.First,the CBAM-GAN network architecture is improved to further enhance the generation ability of the generator.Then,a quality evaluation network is used to automatically pick out high-quality images,which reduces the influence of noisy images on the recognition accuracy.The comparison experiment of image generation on the miniImageNet dataset shows that the image quality generated by Progressive CBAM-GAN has a greater improvement than CBAM-GAN.Finally,adding generated images into the few-shot dataset for training also further improves the accuracy of few-shot recognition.