Research On Scene Image Retrieval Based On Deep Convolution Feature Aggregation

In daily life,people often pay attention to the scene information contained in the image.It is eager to quickly obtain the relevant information(such as name,history,structure,value,etc.)behind the scene by simply taking pictures of the current scene.By processing and coding the target scene image,and combining with certain image matching algorithm and specific scene database,similarity matching retrieval can not only meet the query needs of tourists,make them obtain the corresponding description information behind the scene,but also expand services according to the retrieval results.Therefore,it is of great significance to make full use of the information contained in the scene image and give full play to the important value of the scene image in image retrieval.By analyzing the related problems and key technologies in scene image retrieval,this paper proposes a scene image retrieval technology based on deep convolution feature aggregation.On the basis of PWA algorithm and geographic information filtering,a scene image retrieval algorithm based on salient local feature aggregation and geographic information is proposed.The main research contents involved include:(1)Research on feature aggregation method of deep convolution.The main work is to compare the differences between different "probabilistic proposal" selection and local detector weighted aggregation methods in the "probabilistic proposal" selection stage and local detector weighted aggregation stage,and select the best way to extract and represent the scene image features.(2)The image database for scene image retrieval algorithm testing is constructed,and the geographic information is added to the scene image retrieval algorithm as supplementary information for auxiliary retrieval.A real scene image data set is collected and established.On the basis of the existing scene image database in the laboratory,the geographic information is supplemented,the number of images is increased,the database is improved,and the scalability of the database is ensured,so as to facilitate the subsequent addition of other information of scene images.Finally,the improved algorithm is analyzed and tested by using the constructed database.It provides a test basis for similar algorithms,so that it can be fully applied in the field of deep learning scene image retrieval and recognition and other related research work in the future.(3)A simple scene image retrieval system is developed based on the constructed database and the tested algorithm.Through the system,users can input specific scene images to accurately retrieve the corresponding categories of images and the relevant information behind the images.