Research And Implementation Of Person Re-id Based On Triplet Joint Generative Adversarial Network

In recent years,with the continuous development of artificial intelligence and people's attention to public safety issues,person re-identification(re-id)task has become a hot topic in the field of artificial intelligence.Person re-id task is to judge whether a person captured by multiple cameras that do not have overlapping fields of view is the same one.It depends mainly on the low-level visual features and the high-level invariant attributes extracted from the images rather than the subjective cooperation of the captured person,which thus has a great research value.On the other hand,it has an important role in the application of security as it aims to find and identify the target person accurately and swiftly from massive data.At the same time,the research of person re-id has shown a great significance for the convenience of our everyday life such as album clustering and human-computer interaction.With the successful application of convolutional neural network(CNN)in computer vision,person re-id task has achieved a leap from traditional methods based on handcrafted features to those based on deep learning.However,the person re-id methods based on deep learning often require a large number of high-quality data while data labeling is laborious and costly.Therefore,researchers introduced the idea of generative adversarial network(GAN)on the basis of CNN,and carried out data augmentation by generating new data via GAN.However,in these methods,the performance of person re-id model is improved from a single source,only through the additional data obtained from augmentation.In this paper,two joint models based on representation learning and metric learning are proposed to realize the triplet joint training of the person re-id model and the generator and discriminator of GAN.Through the construction of true-and-false sample pairs,the adversarial learning between the person re-id model and the GAN discriminator can promote and enhance each other in the joint training process.At the same time,the performance of the generator is improved by improving the performance of the discriminator,which makes the generator generate samples of higher quality.These generated samples and the original samples are mixed as the training set of the person re-id model,which further improves the performance of the person re-id model.In this thesis,extensive experiments were carried out on three widely used data sets,namely Market-1501,DukeMTMC and MSMT17,based on representation learning and metric learning,respectively.Both quantitative and qualitative experimental results show that the proposed joint training model can improve the performance of the person re-id model and GAN simultaneously.