Data enhancement pedestrian re-identification method based on generative adversarial network model

Abstract

The invention relates to a data enhancement pedestrian re-identification method based on a generative adversarial network model. The method comprises the following steps: segmenting a mask image of apedestrian in an image by using a Mask-RCNN image segmentation algorithm; training an end-to-end improved star-shaped generative adversarial network in combination with the mask image and the manuallylabeled pedestrian attributes, and generating false training images under any number of cameras from the real pedestrian image under one camera; generating false training images of all camera domainscorresponding to all real images by using the trained improved star generative adversarial network; and sending the real image and the false training image into a pedestrian re-identification model,calculating the distance between the pedestrian images, and completing a pedestrian re-identification function.using Mask-to perform pedestrian re-identification; segmenting a mask image of a pedestrian in the image by using an RCNN image segmentation algorithm; training an end-to-end improved star-shaped generative adversarial network in combination with the mask image and manually labeled pedestrian attributes, and generating false training images under any number of cameras from a real pedestrian image under one camera; using the trained improved star-shaped generative adversarial network to generate false training images of all camera domains corresponding to all real images; and sending the real image and the false training image into a pedestrian re-identification model, calculatingthe distance between the pedestrian images, and completing the pedestrian re-identification function. The method is reasonable in design, more training samples are generated through the generative adversarial network, meanwhile, the generated image background can effectively represent the real scene under the corresponding camera, the robustness and the judgment capability of the pedestrian re-identification model are effectively improved, and the accuracy of pedestrian re-identification is effectively improved.

Description

technical field [0001] The invention belongs to the technical field of computer vision, in particular to a data-enhanced pedestrian re-identification method based on a generative confrontation network model. Background technique [0002] In recent years, security camera networks have spread all over the city, and with it comes the storage of massive video data. However, due to the time-consuming and laborious manual analysis of a large amount of video content, person re-identification technology, as an important part of intelligent video surveillance, has attracted the attention of researchers all over the world. [0003] Pedestrian re-identification technology refers to the specific pedestrian technology that uses computer vision technology to match different scenes in a non-overlapping multi-camera system. Due to the change of the environment where the camera is located and the non-rigid motion of pedestrians, such as brightness changes, pose changes, occlusion, and backg...

AI Technical Summary

Problems solved by technology

However, existing deep learning models still have shortcomings: in actual scenes, the number of images of pedestrians is extremely large, and the scene changes are diverse. However, the size of the existing public datasets is relatively small , and only collect images of pedestrians under a few specific cameras

In addition, due to the relatively small number of cameras used in the collection of existing data sets, the internal environment changes are relatively fixed, and it is difficult to match the real scene

Therefore, the existing deep learning-based pedestrian re-identification models are prone to overfitting in the training process using public datasets.

Images