Research On Few-shot Image Classification Algorithms Based On Data Augmentation And Its Application

Machine learning is an important part of research filed of artificial intelligence that enables machines to exhibit autonomous decision-making and intelligent behavior by learning patterns and rules from training data.In the field of machine learning,large-scale labeled data is usually required,and various algorithm models are used to learn the labeled data,thus obtaining a model that can accurately predict and identify.However,in some application scenarios,obtaining a large amount of training data is not feasible.When faced with these scenarios with scarce training data,the performance of machine learning models will be seriously degraded.To address this issue,researchers have proposed the concept of few-shot learning,which aims to enable machines to quickly learn and make accurate decisions from only a small amount of labeled data.Few-shot image classification is an important application in the field of few-shot learning.For few-shot image classification problems,there are usually three types of methods:metric-based,meta-learning-based,and data augmentation-based.Among them,data augmentation-based methods effectively address the issue of sample shortage in few-shot learning field by designing efficient algorithms to expand the available training sample size.Compared with the other two methods,it has achieved better results in many practical applications.From the perspective of data augmentation,this paper studies the few-shot image classification algorithm,and the main research contents are as follows:(1)To address the problem that few-shot image classification based on feature distribution calibration can not accurately reveal the feature distributions of novel classes,an improved algorithm combining latent space transform and density-based spatial clustering is proposed by this paper.The improved algorithm constrains the feature distribution of the novel classes by means of the latent space transform method,and then simulates the real distribution of the novel classes by means of the density-based spatial clustering method,and finally,the data augmentation of few-shot data is completed by sampling in the simulated distribution of the novel classes as expanded samples.The experimental results show that the proposed improved algorithm can further improve the accuracy of few-shot image classification compared to the traditional feature distribution calibration model.(2)Aiming at the problem that the generative adversarial network does not perform well in the field of few-shot image classification,this paper proposes a few-shot image classification algorithm based on feature distribution calibration model and generative adversarial network.The algorithm uses samples of the base classes similar to the few-shot data to train the discriminator network,and samples from the feature distribution after feature distribution calibration as the noise input for the generator network.This enables the generative adversarial network to generate similar samples with high confidence,thus completing the data augmentation task for few-shot data.Comparative experiments on baseline datasets show that this algorithm improves the performance of few-shot image classification based on generative adversarial network.(3)This article applies few-shot image classification algorithms to the field of rice disease and pest identification,and a rice disease and pest image identification system is realized.Firstly,a few-shot dataset containing 13 common rice diseases and pests is constructed using methods such as image collection,data augmentation,and manual labeling,as well as an external dataset based on disease and pest images of other crops.Then,the system is designed and implemented using the Python-QT framework and the Pytorch deep learning framework for the front-end and back-end.The system integrates multiple few-shot image classification algorithms and features various functions,such as algorithm selection,algorithm training,image upload and recognition,and viewing corresponding disease prevention and control measures.It is capable of completing few-shot image classification tasks on the constructed few-shot dataset.