An image re-identification system and loss function determination method based on label optimization

Abstract

The invention discloses a method for determining an image re-identification loss function based on label optimization, which includes obtaining multiple original labeled pictures, and generating multiple unlabeled pictures through a generative confrontation network; for each labeled picture and each The picture without label carries out feature extraction; Calculate the class centers of the described multiple labeled pictures of multiple classes and the multiple clusters of the multiple pictures without labels and the cluster center of each cluster; Calculate the cluster center and the Euclidean distance of each of the class centers; calculate the probability coefficient of each cluster to the loss function of the plurality of classes according to the Euclidean distance, and obtain the loss function. The present invention also discloses a label-based image reconstruction The recognition system solves the over-fitting phenomenon that is prone to occur when there are not many labeled pictures, and improves the accuracy of re-recognition.

Description

technical field [0001] The invention relates to the technical field of machine learning. More specifically, it relates to an image re-identification system and loss function determination method based on label optimization. Background technique [0002] Today's re-identification technology uses different deep learning network architectures (such as CNN, ResNet, and VGGNet, etc.) and complex algorithms (such as RBM, Adam, and RMSprop, etc.), and with the help of increasingly advanced hardware conditions (GPU), it has achieved High accuracy, but there is one of the easiest ways to improve accuracy by increasing the data set to allow the network to learn more accurate data features. But obtaining labeled data is relatively expensive, so training is performed by increasing the unlabeled data set. [0003] There are many and easy ways to add unlabeled data, such as downloading directly from the Internet, delabeling labeled data, etc. The latest method is to use Generative Adver...

AI Technical Summary

Problems solved by technology

For the unlabeled data set generated by GAN, there are three methods: 1) All in one: treat all unlabeled pictures as a class; 2) Pseudolabel: each unlabeled picture is obtained for all classes Probability, the class with the greatest probability assigns the label of this class to the picture; 3) LSRO (Label Smoothing Regularization for Outliers): The possible source of a label is a linear combination of all other classes, and pictures without labels do not belong to any category , is a linear combination of all classes, and the coefficients are 1 / K, K is the total number of classes with labeled data, in the loss function, each label is a uniform distribution of the log form of each class probability, but, when there When there are too few images in the label, it is prone to overfitting

Images