Research On Methods Of Few-shot Image Classification Based On Deep Metric Networks

Image classification is a fundamental and important task in the field of computer vision.The existing deep learning methods can effectively extract the abstract features of images,so the large sample image classification methods based on deep learning have achieved superior classification accuracy under the condition of sufficient sample size.However,each task in the small sample image classification only contains a small number of labeled samples,which makes the large sample image classification methods unable to apply to the small sample image classification.The existing Few-shot learning methods alleviate the impact of insufficient sample size through deep metric learning,data enhancement and transfer learning,among which deep metric learning is simple and effective,and has achieved excellent performance in many Few-shot image classification tasks.However,the current deep metric-based learning methods still have some problems,such as measurement bias,poor classification of noisy images,and non-aggregation of similar features.To solve the above problems,this thesis studies the Few-shot image classification methods based on deep metric learning and completes the following three tasks:1.A duplex-similarity metric network(FEAT-DS)for Few-shot image classification is proposed.The existing Few-shot image classification methods only consider the distance between the query sample and the class prototype,and do not consider the distance between the class prototype and the query sample,which makes the model produce large measurement deviation.Based on this,this thesis designs a new duplex-similarity metric network for Few-shot image classification.Specifically,based on the Few-shot adaptive embedded converter,a duplex-similarity metric calculation method is constructed by calculating the distance between query samples and class prototypes,which alleviates the metric deviation problem in the calculation.Simulation studies show that this method can calculate the distance between query samples and class prototypes more accurately and identify query sample categories more accurately without increasing any computational cost.2.An anti-noise relation network(AN-RN)for Few-shot image classification is proposed.The existing Few-shot image classification methods pay too much attention to the classification accuracy of clean images,but ignore the classification accuracy of noisy images.To address this issue,a new anti-noise relation network is proposed in this thesis.Specifically,the sample is reconstructed through the Auto Encoder to mitigate the impact of image noise on the model,and an anti-noise relation network loss function is constructed to enhance the robustness of the model.Simulation studies show that this method achieves good performance in both noise-free and noisy settings.Compared with the relation network,the advantages of this method are more obvious with the increase of noise intensity.3.A nearest neighbor metric prototype network(NN-Proto Net)for Few-shot image classification is proposed.Most of the existing Few-shot image classification methods only consider the distance between query samples and support samples,and do not constrain the distance between similar samples in the support set,which leads to the dispersion of sample features.Motivated by this,this thesis proposes a new Few-shot image classification method,the nearest neighbor metric prototype network.Specifically,after calculating the prototype of each category,the proposed method further aggregates the features among the same kind of samples through the nearest neighbor component analysis to make the image features of the same category closer and the image features of different categories farther.Simulation studies show that on multiple Few-shot datasets,the classification accuracy of NN-Proto Net for query images is better than that of the comparative Few-shot image classification methods.