Research On Domain Adaptation And Domain Generalization Methods For Image Recognition

In recent years,the application of image recognition technology in human life is more and more extensive,and the function of image recognition model is more and more powerful.In order to meet different tasks,recognition models need to be trained with a large amount of targeted data.However,due to the phenomenon of "dataset bias",the dataset that used to train models often does not cover all the data distributions that occur in real applications.Therefore,if the models that have been trained by a specific dataset are directly transferred to the actual task,their recognition performance will be significantly reduced.In order to solve the above problems,the paper studies how to improve the transfer performance of image recognition models based on domain adaptation and domain generalization.Based on the original model,a series of improvements have been made,and the specific research results are as follows.1 、 In terms of domain adaptation,aiming at the problem that the performance degradation of the existing advanced models due to the shortcomings such as the insufficiently dispersed distribution between individuals in the compound domain during training,a Markov Open Compound Domain Adaptation model is proposed.Firstly,the Markov process is applied to mix multiple domains together to form a Markov Compound Domain,and an open domain is added to this compound domain to construct a Markov Open Compound Domain.This process ensures that the distribution between individuals that from different domains in the compound domain is sufficiently dispersed.Secondly,in order to solve the problem that the model extracts inadequate image features,a neural network encoder based on Parametric Rectified Linear Unit is proposed.Finally,combining the Markov Open Compound Domain and the new neural network encoder,the Markov Open Compound Domain Adaptation model is trained through the idea of curriculum domain and GAN.2、In the research work on domain generalization,on the one hand,aiming at the problem of uneven distribution of search directions in the process of searching image attributes by existing algorithm models,a Minimum Weighted Random Search algorithm and an "Equal-Sum" judgement method are proposed,which allows the model to appropriately "ignore" those objects that have been selected too many times during the training process.On the other hand,aiming at the problem that the under-utilisation of image attributes in the training process,the complexity of the optimisation combination is increased and the neural network model is improved.Combining the Minimum Weighted Random Search algorithm,the "Equal-Sum" judgement method and the improved network model,three domain generalization improvement models are proposed: the Minimum Weighted Random Search Data Augmentation algorithm,the Minimum Weighted Evolution-Based Search Data Augmentation algorithm and the Minimum Weighted Random and Evolution-Based Search Data Augmentation algorithm.3、The models are trained and tested on publicly available datasets to validate the performance of the models.Aiming at the Markov Open Compound Domain Adaptation models from the research work on domain adaptation,the digit recognition experiments and face recognition experiments are carried out.Three improved models for the research work on domain generalization are implemented for digit recognition experiments and CIFAR-10 classification experiments.Based on the accuracy of the models,a "model transferability " evaluation criterion is proposed to evaluate the models,which reflects the overall performance of the models.The experimental results show that both the domain adaptation model and domain generalization model proposed in this paper can obtain higher recognition performance than the baseline model after being transferred and applied,which verifies the validity of the models.