Research Of Domain Generalizatio Algorithm Based On Deep Learning

Although convolutional neural networks(CNNs)have shown excellent performance at different computer vision tasks,they do not cope well with domain shifts and pose fragile generalization ability.Recent studies find that the domain shift mainly lies in the style or texture variation of images rather than the content.Inspired by this,this paper propose dynamic style transferring and content preserving for domain generalization to overcome the style bias of CNNs.Specifically,this paper design a knowledge-injected attention mechanism to learn adaptive fusion weights and embed the style knowledge of dynamic chosen images in latent space.So the extent of transferred style is controlled,and the contentrelated information is maintained.Further,we devise the content preserving module,which introduces an adversarial structure with the encoder to make the extracted style information more precise.For balancing the adversarial relationship between encoder and auxiliary predictor,we also enforce a consistency loss to empower the style-biased predictor and indirectly boost the encoder’s ability by extending the back-propagation process.The effectiveness of the approach presented in this paper can be observed through ablation experiments on the attentional mechanism of knowledge injection and loss of coherence.To focus on class-relevant regions and preserve the categoryrelevant information,we design the attention consistency loss to guide the learning of model.Furthermore,the bound of generalization error on unseen domain is reduced by enforcing the prediction probability of different classes obey a prior distribution.In this paper,a large number of experiments are carried out on PACS,OfficeHome and Domain Net,and the results of the visualization experiments are presented.Extensive experiments show remarkable performance over the state-ofthe-art methods in the domain generalization.