A Multi-Loss Joint Training Method Preserving the Consistency of Multiple Metric Spaces

Abstract

The invention discloses a multi-loss joint training method that maintains the consistency of multi-metric spaces, which includes the following steps: pedestrian samples are propagated forward through a convolutional neural network to obtain a high-dimensional feature matrix; then converted into pedestrian feature vectors; and the feature vectors are subjected to L2 regularization Operation; the unit eigenvectors are concatenated to form a triplet, and the triplet loss is calculated; the unit eigenvectors are obtained through the Bach Normalization layer to obtain the test vector; then input to the fully connected layer after the network model and the cross entropy loss is calculated forward; calculate its gradient information ; Gradient backpropagation of the loss layer, update the weight parameters; if the model has not yet converged, or the maximum number of iterations has not been reached, repeat the above steps. Without increasing the amount of parameters, using the method of the present invention on each pedestrian re-identification model can significantly improve the effect.

Description

technical field [0001] The invention belongs to the technical field of neural networks, and in particular relates to a multi-loss joint training method for maintaining the consistency of multi-metric spaces. Background technique [0002] As a deep learning model, deep convolutional neural networks have achieved state-of-the-art performance on many computer vision tasks such as image classification, object detection, and object segmentation. In the same way, the deep learning model promotes the rapid development of the field of pedestrian re-identification with the help of a larger amount of parameters and a stronger generalization ability. The pedestrian re-identification model is often used in the field of security inspection. Based on a single pedestrian image, the pedestrian matching operation can be completed for a large number of video images in the hard disk. Thereby, the location information of the pedestrian in different time periods, and the appearance of the pede...

AI Technical Summary

Problems solved by technology

At the same time, the sampling of triplet loss is random, so it will cause the feature vector received by softmax loss to change continuously, making it difficult to converge

Images