Research On Cross-domain Image Retrieval Algorithm Based On Visual Feature

With the rapid development of imaging sensor performance and abundance of sensor types,the number of cross-domain images which are from different imaging carriers,different imaging spectra,and different imaging conditions increase exponentially.To make full use of these digital resources,people often need to integrate different imaging sensors to obtain more comprehensive and rich information.Cross-domain image retrieval,which investigates the problem of searching images across different visual domains,has become one of the research focuses in the field of computer vision.In addition,cross-domain image retrieval is widely applied in many areas,including heterogeneous image registration and fusion,visual positioning and navigation,scene classification,etc.Therefore,there exists important theoretical significance and application value to make a depth research of cross-domain image retrieval.This thesis introduces the research status of cross-domain image retrieval,deeply analyzes the inherent relationship between different visual domain images,focuses on three key problems:cross-domain visual saliency detection,cross-domain feature extraction and description,cross-domain image similarity measurement,implements a cross-domain visual retrieval approach based on visual saliency detection,a novel visual vocabulary translator based cross-domain image retrieval and the co-occurring feature for cross-domain retrieval.The main research contributions are summarized as follows:(1)After analysis the visual saliency of cross-domain image,a cross-domain visual retrieval method based on saliency detection is presented.Firstly,this algorithm sets different values of each super pixel region according to their boundary connectivity and obtains the subject area.Then,image features are optimized by linear classification and each database image is processed in multiple scale.Finally,the similarity degrees between query and each database images are calculated,and the highest scoring image is the retrieval result.This method can effectively reduces the interference of background and other irrelevant regions,and improves the retrieval accuracy.(2)Aimed at the problem of image features from different visual domains vary widely and they cannot be matched directly,a cross-domain image retrieval method based on a novel visual vocabulary translator is proposed.Inspired by language translation,this method exploits a visual vocabulary translator to establish the relationship between differen visual domains.This translator consists of two main modules: one is vocabulary tree which can be regarded as the codebooks in their respective fields,while the other is the index file attached with the leaf node.The index files records the translation relations.Through visual vocabulary translator,cross-domain retrieval can be turned into intra-domain retrieval.And image retrieval across different visual domains can be solved from a new perspective.Experiment results verifies the performance of this algorithm.(3)On the basis of the symbiotic correlation between different visual domains,crossdomain co-occurring feature for cross-domain retrieval is proposed.This approach employs the target consistency between cross-domain images and starts from an overall angle of the query.With the symbiotic relationship between cross-domain visual features,the proposed method constructs the cross-domain co-occurring feature and image retrieval across different visual domain is realized on the basis of co-occurring feature.Extensively experiments show that the cross-domain image retrieval method with co-occurring feature can be successfully applied on visible-infrared image retrieval,and the proposed algorithm effectively enhances the retrieval performance.