Multi-feature extraction network cross-resolution person re-identification based on SR technology

Abstract

In the real world, the resolution of the image that is collected can vary depending on the camera's quality or the change in the distance from the pedestrian. Important information is lost from the low-resolution image. It can be difficult to match Low Resolution (LR) input photographs with High Resolution (HR) gallery images. Thus, we suggest that the super-resolution module and the multi-feature extraction module be improved in order to address the aforementioned issues. To be more precise, the resolution of the low-resolution query image is restored in the first step using an upgraded Super Resolution (SR) model (VDSR-NAM). A two-stream feature extraction network extracts and fuses the features of the LR and SR images in the second stage. The potential of our model has been shown in numerous tests on cross-resolution person re-id datasets. The efficacy of the loss function on our model is concurrently confirmed by ablation experiments on the dataset MLR-VIPER.