Research On Person Re-identification Algorithm Based On Natural Language Description

Person re-identification is based on the person targets appearing in the current surveillance scene,and re-identifies the person shot in the surveillance cameras of different scenes through computer vision technology.However,in actual applications,images or videos of person cannot be fully captured.The natural language description of the person being queried provides more useful information for person re-identification,Therefore,person re-identification based on natural language description becomes research hot spot.Person re-identification based on natural language description belongs to the problem of cross-modal matching.The features of images and natural language description text are obviously heterogeneous.For this purpose,a two-branch extraction network model is constructed,and a truncated attention mechanism for natural The features of language description text are reprocessed to improve its characterization.In the similarity measurement of the features in two modes,a cascade loss function based on relative entropy is designed to realize the similarity measurement.The main work of this thesis is as follows:A two-branch network framework based on natural language description for person rerecognition is proposed to extract features of images and natural language description text respectively.Among them,the features based on the Mobile Net network are used to extract image features;the natural language description text is preprocessed by word embedding,and then the features of the Bi-LSTM network are used to extract the temporal features of the text.In order to improve the representation of natural language description text features,a truncated attention mechanism is proposed to redistribute the weight of text features,that is,to set the threshold to filter the weight of words,highlight the local feature vectors that are obvious,and ignore the low significance.Local features of the text,and finally describe the features of the text as natural language.A cascade loss function based on relative entropy is proposed to measure the feature similarity,which not only effectively uses the person identity label,but also solves the problem of difficult training and design of the loss function.cascade loss function is a fusion of cross-modal matching loss function based on relative entropy and single-mode multi-class loss function based on relative entropy.Cross-modal matching uses relative entropy to reduce the difference between the probability distribution of image and text feature matching results and the probability distribution of actual labels.Single-modal multi-classification,which reduces the difference between image and text multi-classification probability distributions by relative entropy.This thesis conducts experiments on the only CUHK-PEDES dataset in the field of person re-identification based on natural language description.The experimental results show that the accuracy of the algorithm Top-1 is 50.19%,the accuracy of Top-5 is 72.50%,and the accuracy of Top-10 is 80.77%.Compared with the current mainstream algorithms,this algorithm is based on ensuring that the network structure is simple and flexible.It can effectively solve the problem of person re-identification based on natural language description.