Landmark Images Recognition With Deep Features

Landmark recognition will be widely used in a lot of applications,the most effective method is to use image retrieval to query similar image's location.There are still several problems unsolved in these deep feature retrieval methods:(1)Most models use a single feature.Only global or local features are used in CNN method.How to effectively fuse the two features is still a hard problem today.(2)As the limited of landmark detection datasets,landmark recognition doesn't have the process of detection and recognition,so the accuracy of landmark recognition will be reduced.(3)The standard cross entropy loss function in the classification training does not perform well when the number of categories is far greater than the embedded feature dimension,and the triplet loss is difficult to converge quickly because of the efficiency of triplet mining.In view of the above problems,this subjects have finished the following works:(1)Explore the model structure of gem pooling and attention mechanism.(2)Study the characteristics of comparative loss and its applicable conditions.Give a scientific selection range of training parameters.Combine the model structure and strategy to train the global feature extraction model.(3)Study the Additive Angular Margin Loss(ArcFace loss)and analyze the selection conditions of angle domain parameters and describes the characteristics of this loss.Based on the global feature extraction model,an efficient model of landmark recognition(ArcMF)with multi feature fusion is proposed,which uses weak supervision training.In this project,experiments show that ArcFace loss can effectively improve the effect of the model and ArcFace loss significantly increase the inter-class distance and reduce the intra-class distance.The ArcMF model has achieved better results in the Paris and Oxford building datasets,and the effectiveness of this model is verified in the actual scene.