Research On Image Recognition Method Based On Information Augmentation And Transfer Learning

The field of computer vision has long been known for its frequent and significant research into image recognition.With the emergence of new technologies and continuous iterative updates,the application and development of image recognition has advanced and achieved major breakthroughs.However,these breakthroughs often rely on the image recognition technology based on deep learning,which not only needs to label a large number of images to train the recognition model,but also needs to assume that the image data of the train set and the test set obey the same probability distribution.However,in actual scenarios,due to the problem of domain shift in datasets,the performance will be significantly reduced when the image recognition model of a specific dataset is trained and generalized to other test datasets.The purpose of this paper is to explore how to effectively simulate the target domain to improve the robustness of the model and complete cross-domain image recognition tasks through information augmentation and transfer learning techniques.The paper adopts two methods combining transfer learning: one is to enhance the source domain,and the other is to generate data to improve the accuracy of image recognition.1.With changing image attributes to achieve data augmentation and assist in target domain recognition tasks,an image recognition method based on data augmentation and transfer learning was studied.Firstly,in response to the problem of high search randomness in existing models for searching image attributes,an Average Probability Search Algorithm is proposed,which can dynamically balance the probability of each image attribute being searched.Secondly,a Parallel Genetic Algorithm is proposed to address the problem of low diversity of attribute combinations in the process of constructing image attribute combinations in the model,which increases the diversity of image attribute combinations.Thirdly,in order to fully learn the key Semantic information of the enhanced image,a multi-branch convolutional residual network is designed.Finally,combining the Average Probability Search Algorithm,Parallel Genetic Algorithm,and multi-branch convolutional residual network,three improved and optimized algorithm models were proposed: Average Probability Search Data Augmentation,Parallel Genetic Search Data Augmentation,and Average Probability Parallel Genetic Search Data Augmentation.2.With generating images different from the source domain to simulate the target domain and assist in the recognition task of the target domain,an image recognition method based on data generation and transfer learning was studied.Firstly,in response to the problem of insufficient diversity and availability of generated images in existing models,a multi-branch generation network was constructed by obtaining multi-scale feature maps of different fields of view through convolutional encoders.Secondly,feature diversification is achieved through self-attention normalization layer,and different distributed images are synthesized through decoder,so as to achieve the effect of simulating data distribution in the target domain.Then,contrastive learning and two-way content consistency strategy are introduced to ensure the effectiveness of generated images,while enhancing information exchange and learning invariant representations between samples to improve the ability of the model to identify features.Finally,combined with multi-branch generation network,backbone network,contrastive learning and two-way content consistency strategy,an image recognition model based on data generation and transfer learning is designed.3.Experiments on common transfer learning image recognition datasets are conducted to evaluate the generalization ability of the model in order to achieve better recognition results in this paper.The experimental results indicate that in both digital recognition experiments and CIFAR-10 classification experiments,two kinds of image recognition models based on information augmentation and transfer learning,is superior to the benchmark model,thus confirming its effectiveness.