Instance Search Based On Visual Content

Instance search is a task proposed by international computer vision competition TRECVID: Given a specific i nstance(people,o bjects,s cenes),s earch t he v ideo clips or images containing this instance in dataset.Instance search is a highly practical task,which is of great significance in animal and plant identification,object search,and surveillance security.Not only task goals but also implementation methods,instance search is obviously different from traditional content-based image search and near-duplicate search.Recently,instance search based on visual content has made effective progress.However,there are still many problems:(1)The original mature method of spatial consistency verification requires that the query need be compared with all images in the database,and all features are involved in the matching.Therefore,time cost of the process is large.(2)The popular deep feature methods do not make good use of characteristics of instance search.They are just suitable for general image retrieval problem.In this thesis,we propose two methods to address these two issues.The first method focuses on improving the efficiency of the spatial v erification.In order to fit the conditions and application scenarios of mobile devices,we propose an instance search algorithm suitable for mobile devices.We firstly mine images that are similar to the query on mobile devices.Then we propose three criteria to select representative and concise visual words.Finally,a simple and effective spatial verification method is used to compute the similarities between query and images.The second method focuses on combining deep features with the characteristics of instance search.We generate candidate regions for each image and extract regional features.Then,we use the region features for feature refinement and spatial verification to generate a similarity network of im ages.After that,the method extracts effective and refined core features based on the community detection.The features of the query directly matches with the core features,which can improve efficiency and performance.To prove the validity of our methods,we conducted experiments on the Oxford Building dataset.