Applying A Newglobal Loss Function With Fused Multipe Loss Function In People Re-identification Neurals Networks

With the rapid development of automation monitoring,researcher has more and more notice pedestrian re-identification in recent years.The number of video images in the monitoring system is massive and quasi-big data,which not only brings great pressure on image transmission and image storage,but also on the monitoring system and challenging atable and reliable operation.Gaining valid data from large amounts of data manually is very inefficient.As for computers,even accurately extracting pedestrian information from a rich image is quite.Therefore,the topic of how to make the computer simulate human vision and realize the extraction of pedestrian information has gained more and more attention fr-om researchers.In training deep convolutional neural network models to perform the task of person re-identification,loss function is critical to the final performance.In this paper,we propose the use of a new loss function that represent the general distribute of data to reduce the classification error.By take in count of the mean and variety of similarity measurement between the data of same label and different label,we can reduce the false positive and false negative classification error.Therefore the model can gain a better capability.Furthermore,we combine several establish loss function with this new loss function to get a proposed jointly training network model.Our network can learns similarity measurement between the data of same label and different label at the same time and each bit of data in datasets can contribute to final result.The main work of this paper and its main contributions are the following three aspects:1.In training of the convolutional neural network for the people re-identification task,the loss function between the pairs of images compares only one pair of images at a time,and the global distribution of the features acquired from the image data is lost in the process.It is proposed to use a new loss function using global data distribution to reduce classification errors and use th lost data.2.The identification model and the verification model commonly used in the pedestrian re-identification problem have their advantages and limitations,it is proposed to integrate them and jointly perform the task of people re-identification to overcome the shortcomings of the two and develop their respective advantages.3.In the construction of the convolutional neural network framework,a siamese neural network is used as the basic structure,and VGG16 and Resnet50 are used as the image feature extraction network,and the effects of a series of hyperparameters on the experimental results are experimentally studied.,indicating that the deep convolutional neural network can produce very competitive accuracy in the re-identification compared to the prior art through this new loss function.we show that the our network with this new loss function can yields very competitive re-id accuracy compared with state-of-the-arts one.