Feature Enhancement Based Algorithm For Few-shot Image Classification

Few-shot learning(FSL)is a machine learning technique that aims to train a model by using a small number of examples in high-dimensional data.In traditional machine learning,a large amount of data is usually required to train an accurate model.However,in practical scenarios,it may be very difficult to obtain a sufficient amount of data because data acquisition involves security,ethics,resources and costs,so few-shot learning has become a topic of much concern.Few-shot learning only needs a small number of samples to learn quickly,which greatly reduces the cost of manpower,material resources,financial resources and time in model training,and has a wide range of application prospects in the field of data shortage.However,at the same time,due to the limited number of labeled samples in few-shot learning,in order to avoid overfitting,a backbone network with shallow depth and narrow width is usually selected.However,such a backbone network can only extract shallow features of images,and the accuracy and adequacy of feature extraction are not enough.Therefore,this paper proposes several feature enhancement methods to improve the feature extraction ability of few-shot models.The main innovative work of this paper is as follows:(1)Object region localization method based on instance segmentation.Due to the limited feature extraction ability of the backbone network,the interference of background and pseudo-target noise in the image on the prediction accuracy of the model is more obvious in small sample learning.Therefore,this paper proposes to obtain the positioning image containing only the target object through instance segmentation,and accurately eliminate the interference of background and spurious target noise on the accuracy of model prediction.(2)Multi-level distribution extraction method based on graph neural network.A large number of experiments show that the deep network integrates the shallow,middle and high-level features of the image,and the "hierarchy" of features gradually becomes richer by stacking the number of network layers.Therefore,this paper proposes to use the propagation characteristics of graphs to obtain the distribution of image features,and increase the depth of feature extraction layer by layer by stacking multiple graph neural networks in series to obtain more semantic information.(3)Frequency domain feature extraction method based on discrete wavelet transform.Discrete wavelet transform can decompose an image into subbands with different frequency information and spatial details.In this study,the four sub-bands obtained by two-dimensional discrete wavelet transform of the image are spliced on the high and wide channels,and input into the backbone network together with the original image for training,so that the network can learn more comprehensive and diverse feature information.(4)Unsupervised learning method based on k-means clustering.Unsupervised kmeans clustering algorithm does not need labels.It calculates the distance between image samples to measure the similarity of samples,divides similar images into the same cluster,and achieves clustering by maximizing the distance between clusters and minimizing the distance within clusters.It is simple,easy to implement and has strong interpretability of clustering results,which can well understand the data characteristics and rules in image classification tasks.In summary,this paper proposes four improved methods based on feature enhancement based on the problem of inaccurate and inadequate feature extraction in the small sample task,and finally uses the results obtained by the four methods through the voting mechanism to obtain the final output prediction results.After experimental verification,the model based on a series of improved methods proposed in this paper has higher few-shot image classification accuracy than the baseline: the classification accuracy of the model is 8.40% higher than the baseline in the 5-way 1-shot setting,and2.96% higher than the baseline in the 5-way 5-shot setting.Meanwhile,under the setting of the same dataset and backbone network,the proposed model also outperforms most existing few-shot classification models.