Deep Image Recognition Research For Small Sample Based On Metrics

Currently,convolutional neural networks(CNN)have shown their effectiveness in many computer vision tasks,such as image classification and scene recognition;however,training CNNs on these tasks requires a large amount of training data.Compared with CNN,human beings can learn new concepts from only a small number of samples.Nowadays,in many scenes,such as medical images,frequent animal and plant identification,etc.,the problem of small sample sizes makes it difficult to train CNNs.Therefore,it is of great value to study how to achieve better learning with a small number of samples.Most existing few-shot learning methods mainly use first-order statistics for conceptual representation.However,in the era of manually annotated features,second-order statistics,especially regional covariance descriptors,have been shown to be effective for visual recognition tasks.However,in the existing work,while training the classifier,there is no learning component involved in the calculation of the covariance descriptor,and this article attempts to learn the second-order statistics from the image.To this end,this thesis proposes a covariance metric network,which uses second-order statistics to improve few-shot learning.Experimental results on the mini Imagenet dataset and a fine-grained benchmark dataset CUB-Birds show that the covariance metric network proposed in this paper achieves competitive results compared to previous techniques.The main goal of few-shot learning is to learn a classifier with good generalization performance with a small sample size.But with less training data,it is difficult to train a feature extractor with good performance to extract accurate sample features.In addition,the small sample classification based on the metric method also has room for improvement in the feature embedding strategy.It needs to further improve its versatility and transferability to handle invisible class samples.Aiming at these problems,this paper has conducted in-depth research on data augmentation,feature embedding and feature measurement.A star triplet metric network is proposed.In addition,an effective saliency guidance data illusion network is used to enhance training data.The experimental results on the mini Imagenet dataset and three fine-grained benchmark datasets Stanford Dogs,Stanford Cars,and CUB-Birds show that the star triplet metric network proposed in this paper achieves competitive results compared to previous techniques.