Research And Implementation On Image Classification Algorithm Based On Domain Generalization

In recent years,significant progress has been made in image classification algorithms,primarily applied to practical scenarios such as facial recognition,autonomous driving,and medical diagnosis.However,domain shift is a common problem in image classification tasks,where the classification model’s accuracy in testing data will significantly decrease due to differences in data distribution caused by factors such as lighting,image style,and shooting angle between the training and testing data.This greatly limits the application of image classification algorithms to a broader range of scenarios.Numerous studies typically employ domain adaptation methods to overcome this issue,which mainly involve reducing the distribution difference between the source and target domains,utilizing source domain information to achieve target domain learning.However,in specific scenarios when target domain data is unavailable or unknown,this method becomes inapplicable.Domain generalization emphasizes that the learned model from the source domain can generalize to any unknown domain and achieve good performance in image classification tasks with unknown testing data.This issue is more applicable to real-life scenarios,but it also poses challenges and,therefore,holds significant research value.As data determines the model’s upper limit,data augmentation methods are an important approach to solving domain generalization problems in image classification.The specific research content of this article is as follows:Firstly,this article proposes a feature augmentation method based on Haar wavelet transform,which addresses the issue of semantic distortion in original features caused by existing feature enhancement methods during domain generalization,thus affecting the classification accuracy of the classification model.Utilizing the property of lossless reconstruction of signals through Haar wavelet transform,we decompose the features into high-frequency parts containing semantic information and low-frequency parts containing style information;then,we apply stylization only on the low-frequency parts while keeping the high-frequency parts unchanged,followed by feature reconstruction using the mirror operation of wavelet transform.The advantage of this method is that it enriches the diversity of training data while fully retaining the semantic information of the original features,thus improving the model’s generalization ability.Additionally,we design a consistency loss function to minimize the prediction difference between original features and stylized features,further improving the model’s generalization ability.Our method achieves the classification accuracy of 84.90%and 66.41%on two mainstream datasets,PACS and Office-Home,respectively,which is a significant improvement compared to most existing domain generalization image classification methods,thus verifying the effectiveness of our proposed method.Secondly,this paper proposes a data augmentation method based on WCT style transfer to increase the diversity of style patterns in the training samples,which helps adjust the bias of the classification model and allows it to focus more on features such as shape that are less affected by domain shift,thereby improving the generalization ability of the model.Moreover,to address the issue of content distortion in images generated by the AdaINbased style transfer data augmentation method for domain generalization,we introduce the WCT operation to preserve the content details of the generated stylized samples more completely and with better quality,which benefits the accuracy of the classification model.Extensive experiments on mainstream domain generalization datasets demonstrate the competitiveness of our method compared to most existing domain generalization image classification methods.Thirdly,leveraging our proposed method,this paper devises and deploys a prototype system for image classification based on the domain generalization image classification algorithm.This system is capable of providing users with satisfactory classification performance even when dealing with images from unknown domains.Moreover,users can employ the prototype system to generate customized high-quality stylized images to augment the training dataset.