Research On Person Re-identification Algorithm Based On Middle-Level Features And Visual Saliency

Person re-identification,which aims to compare a person of interest as seen in a"probe" camera view to a "gallery" of candidates captured from a camera to find objects that belong to the same target entity as those interested in the"probe"camera view.Existing methods mostly concentrate on two aspects:one concentrated on improving low-level features to represent the appearance of a person;the other focused on distance metrics to reduce the gap of the different features of a person.Various methods utilize the global information as the feature descriptors,which neglect the details in an image and may have a low accuracy of person re-identification.On the other hand,not much attention has been paid to suppress the background information,which has an influence on person re-identification to some extent.In addition,in the field of image saliency research,the co-saliency detection of group images is often based on the saliency of single image.Many existing methods mostly deal with the two separately.Few people used a unified model to solve the problem of single image saliency and co-saliency among the group images simultaneously.Being aware of these problems,this paper first proposes a unified framework of saliency detection algorithm,focusing on single image saliency and co-saliency among the group images at the same time,then applies the saliency detection method to the person discrimination network based on the visual saliency and sheds light on a combination framework by fusing the attribute-identity discrimination network with the person discrimination network based on the visual saliency to improve the accuracy of person re-identification by optimizing feature description.We use the attribute-identity discrimination network,which aims at the incompleteness of the global features in the existing methods and combines attribute features with entity features to solve the problem of lower recognition accuracy caused by overly similar pedestrian features.The attribute-identity discrimination network makes use of the comprehensiveness of entity features and the locality of attribute features,respectively.By combining this two kinds of information,the local cue information can make up for the missing part of the global information obtained from the entity,and make the two complement each other,so as to get the feature descriptor which can more comprehensively represent the pedestrian characteristics.Through a lot of experiments,we can know that the middle-level attribute features have more advantages in the accuracy of recognition.We propose a person discrimination network based on visual saliency.In this network,image saliency detection method based on multi-instance learning is applied,which pays more attention to the main object in the foreground of a pedestrian image,and restrains the background information that may interfere with the recognition in the pedestrian image.Before the experiment,the visual saliency was used to process the image of the input network.On the premise of strengthening the pedestrian entity and weakening the background interference,the images are input into the entity discriminant network,and then combined with the more comprehensive pedestrian feature descriptors based on the attribute-identity discriminant network.After that,we can get more accurate and more representative pedestrian feature descriptors.A large number of experiments have helped this paper to prove the important role of visual saliency in the person re-identification.To a certain extent,it makes entity discrimination network have better robustness.