Person Re-Identification Based On Local Feature And Data Augmentation

Pedestrian re-identification(Re ID)refers to the study of the identification of pedestrian images in different environments.There is a wide range of applications for judging the identity of pedestrians in another set of images from existing pedestrian image information.However,from the current situation,existing pedestrian reidentification techniques still have many problems in handling difficult samples.In terms of the adaptability of the model,some environmental factors within the image,such as lighting,viewing angle,etc.,can greatly interfere with the accuracy of the identification.In addition,there are problems with model migration,insufficient data samples,etc.Today’s research on pedestrian re-identification revolves around both feature representation and metric learning.Feature representations aim to explore robust,environmentally uninterrupted image features,while metric learning is focused on studying distance metrics that enable efficient differentiation of different classes of samples.This paper examines this task in the following ways.(1)A network model architecture NGAN-20 for data enhancement is designed,which improves on the traditional generative adversarial network model in order to match the task feature of pedestrian re-identification.The network can generate samples based on different perspectives,environments,and other conditions,thus solving to some extent the current problem of generally small pedestrian image data sets,while experimentally verifying the effectiveness of this network architecture.(2)A convolutional neural network FN-20 was designed that incorporates pedestrian multiscale features,taking into account both global and local human features to extract more discriminating features for pedestrian reidentification tasks.In addition the whole network combines cross-entropy loss and hard-sample ternary loss for training learning to synthesize an end-to-end pedestrian re-identification feature extraction network under the supervision of a multiple loss function.Experimentally,it is proved that the combination of ternary loss and cross-entropy loss of difficult samples improves the network’s ability to discriminate pedestrian features;the global and local features of the network have a certain complementarity,and the performance of the network is improved after fusion.(3)In response to the widespread misalignment of images in pedestrian data sets,this paper combines a spatial transformation network with the FN-20 architecture to implement an end-to-end network model architecture,and experimentally verifies the effect of each part on model performance.