Transfer Learning Algorithm To Person Re-Identification

With the rapid development of intelligent monitoring equipment and deep learning technology,multi-camera joint monitoring has become a hotspot research.The deep learning person re-identification technology for retrieving designated pedestrians in multiple non-overlapping areas has become an important research direction of multi-camera joint monitoring,and has broad application value and research significance.In the actual application scenarios of deep learning person re-identification technology,affected by complex environmental factors,deep learning networks trained with data sets marked in a single environment often lack adaptability and deployability when used across environments.At the same time,the high labeling cost limits the possibility of marked training in the new environment.It has become a practical problem that has been widely concerned about searching for the re-identification transfer learning algorithm that can apply the existing model to the new environment without labels.The paper focuses on the transfer learning algorithm for person re-identification.With the data generation transfer learning algorithm based on SPGAN network,the style transfer learning method for data generation is improved.First,this paper proposes a person image transfer learning method using semantic segmentation masks.By using the mask generated by the semantic segmentation network to distinguish the pedestrian target area from the background area in the person image,and combining the regression loss function to restrict the identity information of the pedestrian target area,the style transfer network is controlled to keep the person identity information as possible as we can.Transfer learning in the style of cross-domain data sets.After that,the migration data set obtained by style migration learning is used for feature re-learning on the person re-identification network to improve the performance of the person re-identification task in the target domain to complete the goal of person re-identification transfer learning.Subsequently,in order to solve the problem of wrong color translation of pedestrian torso and other components in the method to further improve the performance of the style transfer network,the paper also proposes a person image transfer learning method that uses local features instead of global features.This method first divides the pedestrian depth features generated by the SPGAN network into local features,and then separately performs feature transfer.Finally,the transfered local features are restored to generate a transfer image after stitching and fusion.The method can reduce the color translation errors of pedestrian parts in the images generated by the style transfer network,and can be used in conjunction with the semantic segmentation mask transfer learning method to further improve the performance of person re-identification transfer learning tasks.Methods Transfer learning experiments were conducted between the DukeMTMC-reID dataset and the Market-1501 dataset to verify the effectiveness of the method.